├── .github
└── workflows
│ └── publish.yml
├── .gitignore
├── LICENSE
├── README.md
├── __init__.py
├── core
├── __init__.py
├── pipeline.py
├── recaption.py
└── transformer.py
├── dsd_imports.py
├── dsd_nodes.py
├── examples
├── example_workflow.json
└── workflow.png
├── pyproject.toml
├── requirements.txt
├── utils.py
└── web
└── js
└── showEnhancedPrompt.js
/.github/workflows/publish.yml:
--------------------------------------------------------------------------------
1 | name: Publish to Comfy registry
2 | on:
3 | workflow_dispatch:
4 | push:
5 | branches:
6 | - main
7 | - master
8 | paths:
9 | - "pyproject.toml"
10 |
11 | permissions:
12 | issues: write
13 |
14 | jobs:
15 | publish-node:
16 | name: Publish Custom Node to registry
17 | runs-on: ubuntu-latest
18 | if: ${{ github.repository_owner == 'irreveloper' }}
19 | steps:
20 | - name: Check out code
21 | uses: actions/checkout@v4
22 | with:
23 | submodules: true
24 | - name: Publish Custom Node
25 | uses: Comfy-Org/publish-node-action@v1
26 | with:
27 | ## Add your own personal access token to your Github Repository secrets and reference it here.
28 | personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
29 |
--------------------------------------------------------------------------------
/.gitignore:
--------------------------------------------------------------------------------
1 | # Python
2 | __pycache__/
3 | *.py[cod]
4 | *$py.class
5 | *.so
6 | .Python
7 | build/
8 | develop-eggs/
9 | dist/
10 | downloads/
11 | eggs/
12 | .eggs/
13 | lib/
14 | lib64/
15 | parts/
16 | sdist/
17 | var/
18 | wheels/
19 | *.egg-info/
20 | .installed.cfg
21 | *.egg
22 |
23 | # Virtual Environment
24 | venv/
25 | env/
26 | ENV/
27 |
28 | # IDE specific files
29 | .idea/
30 | .vscode/
31 | *.swp
32 | *.swo
33 |
34 | # OS specific files
35 | .DS_Store
36 | .DS_Store?
37 | ._*
38 | .Spotlight-V100
39 | .Trashes
40 | ehthumbs.db
41 | Thumbs.db
42 |
43 | # Jupyter Notebook
44 | .ipynb_checkpoints
45 |
46 | # Logs
47 | logs/
48 | *.log
49 |
50 | # Local configuration
51 | .env
52 | .env.local
53 | .env.development.local
54 | .env.test.local
55 | .env.production.local
56 |
57 | # Distribution / packaging
58 | *.manifest
59 | *.spec
60 |
61 | # Unit test / coverage reports
62 | htmlcov/
63 | .tox/
64 | .coverage
65 | .coverage.*
66 | .cache
67 | nosetests.xml
68 | coverage.xml
69 | *.cover
70 | .hypothesis/
71 |
72 | .DS_Store
73 | .__MACOSX
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 |
635 | Copyright (C)
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | Copyright (C)
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # ComfyUI-DSD
2 |
3 | An Unofficial ComfyUI custom node package that integrates [Diffusion Self-Distillation (DSD)](https://github.com/primecai/diffusion-self-distillation) for zero-shot customized image generation.
4 |
5 | DSD is a model for subject-preserving image generation that allows you to create images of a specific subject in novel contexts without per-instance tuning.
6 |
7 | ## Features
8 |
9 | - Subject-preserving image generation using DSD model
10 | - Gemini API prompt enhancement
11 | - Direct model download from Hugging Face
12 | - Fine-grained control over generation parameters
13 | - Multiple image resizing options
14 |
15 | ## Installation
16 |
17 | 1. Clone this repository into your ComfyUI custom_nodes folder:
18 |
19 | ```bash
20 | cd ComfyUI/custom_nodes
21 | git clone https://github.com/irreveloper/ComfyUI-DSD.git
22 | ```
23 |
24 | 2. Install the required dependencies:
25 |
26 | ```bash
27 | pip install -r requirements.txt
28 | ```
29 |
30 | 3. Get the model files (two options):
31 | - **Option 1**: Use the `DSD Model Downloader` node in ComfyUI to automatically download the model
32 | - **Option 2**: Download manually from [Hugging Face](https://huggingface.co/primecai/dsd_model) or [Google Drive](https://drive.google.com/drive/folders/1VStt7J2whm5RRloa4NK1hGTHuS9WiTfO?usp=sharing)
33 |
34 | The model files will be stored in:
35 | - `ComfyUI/models/dsd_model/transformer/` (for transformer files)
36 | - `ComfyUI/models/dsd_model/pytorch_lora_weights.safetensors` (for LoRA file)
37 |
38 | 4. Restart ComfyUI
39 |
40 | ## Available Nodes
41 |
42 | 1. **DSD Model Downloader**: Automatically downloads the model from Hugging Face
43 | - Supports downloading from custom repositories with the `repo_id` parameter
44 | - Includes options for model precision (bfloat16, float16, float32)
45 | - Provides memory optimization options (low_cpu_mem_usage, model_cpu_offload, sequential_cpu_offload)
46 | - Optional Hugging Face token support via parameter or HF_TOKEN environment variable
47 |
48 | 2. **DSD Model Loader**: Loads a pre-downloaded model
49 | - Supports custom model and LoRA paths
50 | - Multiple precision options (bfloat16, float16, float32)
51 | - Memory optimization options for different hardware configurations
52 |
53 | 3. **DSD Model Selector**: Helps select models from local directories
54 | - Automatically finds models in the default ComfyUI model paths
55 | - Verifies model existence and provides appropriate warnings
56 |
57 | 4. **DSD Gemini Prompt Enhancer**: Uses Google's Gemini API to enhance prompts for better image generation results
58 | - The API key can be provided in two ways:
59 | - As an input parameter to the node (not recommended for sharing workflows)
60 | - Through the `GEMINI_API_KEY` environment variable (strongly recommended)
61 | - Analyzes both the input image and text prompt to generate improved prompts
62 |
63 | Note: To use the enhanced prompts, connect this node's output to the DSD Image Generator's prompt input and enable the `use_gemini_prompt` option. If no API key is provided, the original prompt will be used.
64 |
65 | 5. **DSD Image Generator**: Generates images with the DSD model
66 | - Supports detailed parameter control:
67 | - Guidance scale (overall, image-specific, and text-specific)
68 | - Inference steps
69 | - Resolution control
70 | - Seed control (0 for random seed)
71 | - Returns both the generated image and the reference image
72 | - Displays progress during generation
73 |
74 | 6. **DSD Resize Selector**: Provides flexible image resizing options for the DSD Image Generator:
75 | - **resize_and_center_crop**: Resizes and center crops the image (default behavior)
76 | - **center_crop**: Simple center crop and resize
77 | - **pad**: Preserves aspect ratio and adds padding to reach target size
78 | - **fit**: Resizes to target dimensions without preserving aspect ratio
79 | - Additional customization:
80 | - Interpolation method (LANCZOS, BICUBIC, BILINEAR, NEAREST)
81 | - Padding color (RGB values for pad mode)
82 |
83 | ## Basic Workflow
84 |
85 | 
86 |
87 | ## Advanced Usage
88 |
89 | ### Memory Optimization
90 |
91 | The DSD model can be memory-intensive. Several options are available to optimize memory usage:
92 |
93 | - **Precision**: Use `bfloat16` (default) for the best balance of speed and memory usage
94 | - **CPU Offloading**: Enable `model_cpu_offload` or `sequential_cpu_offload` for systems with limited VRAM
95 | - **Resolution**: Lower resolution and fewer inference steps can significantly reduce memory requirements
96 |
97 | ### Gemini API Integration
98 |
99 | For optimal results with the Gemini API:
100 | 1. Obtain a Gemini API key from Google AI Studio
101 | 2. Set it as an environment variable: `GEMINI_API_KEY=your_key_here`
102 | 3. Connect the DSD Gemini Prompt Enhancer to your workflow
103 | 4. Enable `use_gemini_prompt` on the DSD Image Generator
104 |
105 | ### Custom Model Loading
106 |
107 | If you have custom DSD models or want to use a different repository:
108 | 1. Use the DSD Model Downloader with a custom `repo_id`
109 | 2. Or manually download the model files and use DSD Model Loader with custom paths
110 |
111 | ## Troubleshooting
112 |
113 | - **Memory Issues**: Try reducing precision (use bfloat16), lower resolution, or fewer steps
114 | - **Gemini API**: Ensure you have a valid API key (can be set via GEMINI_API_KEY environment variable)
115 | - **Model Loading**: If you see errors, try using the Model Downloader node to re-download files
116 | - **Import Errors**: Make sure all dependencies are installed correctly
117 | - **CUDA Errors**: If you encounter CUDA out-of-memory errors, try enabling CPU offloading options
118 |
119 | ## Examples
120 |
121 | Check the `examples` directory for sample workflows.
--------------------------------------------------------------------------------
/__init__.py:
--------------------------------------------------------------------------------
1 | """
2 | Diffusion Self-Distillation ComfyUI nodes
3 | """
4 |
5 | from .dsd_nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
6 |
7 | WEB_DIRECTORY = "./web"
8 |
9 | __all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS", "WEB_DIRECTORY"]
--------------------------------------------------------------------------------
/core/__init__.py:
--------------------------------------------------------------------------------
1 | """
2 | Core components of the Diffusion Self-Distillation (DSD) model.
3 | """
4 |
5 | from .pipeline import FluxConditionalPipeline
6 | from .transformer import FluxTransformer2DConditionalModel
7 | from .recaption import enhance_prompt
8 |
9 | __all__ = ["FluxConditionalPipeline", "FluxTransformer2DConditionalModel", "enhance_prompt"]
--------------------------------------------------------------------------------
/core/pipeline.py:
--------------------------------------------------------------------------------
1 | # Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
2 | #
3 | # Licensed under the Apache License, Version 2.0 (the "License");
4 | # you may not use this file except in compliance with the License.
5 | # You may obtain a copy of the License at
6 | #
7 | # http://www.apache.org/licenses/LICENSE-2.0
8 | #
9 | # Unless required by applicable law or agreed to in writing, software
10 | # distributed under the License is distributed on an "AS IS" BASIS,
11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 | # See the License for the specific language governing permissions and
13 | # limitations under the License.
14 |
15 | import inspect
16 | from typing import Any, Callable, Dict, List, Optional, Union
17 |
18 | import numpy as np
19 | import torch
20 | from transformers import (
21 | CLIPTextModel,
22 | CLIPTokenizer,
23 | T5EncoderModel,
24 | T5TokenizerFast,
25 | )
26 |
27 | from diffusers.image_processor import VaeImageProcessor
28 | from diffusers.loaders import SD3LoraLoaderMixin
29 | from diffusers.models.autoencoders import AutoencoderKL
30 | # from diffusers.models.transformers import FluxTransformer2DModel
31 | from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
32 | from diffusers.utils import (
33 | USE_PEFT_BACKEND,
34 | is_torch_xla_available,
35 | logging,
36 | replace_example_docstring,
37 | scale_lora_layers,
38 | unscale_lora_layers,
39 | )
40 | from diffusers.utils.torch_utils import randn_tensor
41 | from diffusers.pipelines.pipeline_utils import DiffusionPipeline
42 | from .recaption import enhance_prompt
43 | from .transformer import FluxTransformer2DConditionalModel
44 |
45 |
46 | if is_torch_xla_available():
47 | import torch_xla.core.xla_model as xm
48 |
49 | XLA_AVAILABLE = True
50 | else:
51 | XLA_AVAILABLE = False
52 |
53 |
54 | logger = logging.get_logger(__name__) # pylint: disable=invalid-name
55 |
56 | EXAMPLE_DOC_STRING = """
57 | Examples:
58 | ```py
59 | >>> import torch
60 | >>> from diffusers import FluxPipeline
61 |
62 | >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
63 | >>> pipe.to("cuda")
64 | >>> prompt = "A cat holding a sign that says hello world"
65 | >>> # Depending on the variant being used, the pipeline call will slightly vary.
66 | >>> # Refer to the pipeline documentation for more details.
67 | >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]
68 | >>> image.save("flux.png")
69 | ```
70 | """
71 |
72 |
73 | def calculate_shift(
74 | image_seq_len,
75 | base_seq_len: int = 256,
76 | max_seq_len: int = 4096,
77 | base_shift: float = 0.5,
78 | max_shift: float = 1.16,
79 | ):
80 | m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
81 | b = base_shift - m * base_seq_len
82 | mu = image_seq_len * m + b
83 | return mu
84 |
85 |
86 | # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
87 | def retrieve_timesteps(
88 | scheduler,
89 | num_inference_steps: Optional[int] = None,
90 | device: Optional[Union[str, torch.device]] = None,
91 | timesteps: Optional[List[int]] = None,
92 | sigmas: Optional[List[float]] = None,
93 | **kwargs,
94 | ):
95 | """
96 | Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
97 | custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
98 |
99 | Args:
100 | scheduler (`SchedulerMixin`):
101 | The scheduler to get timesteps from.
102 | num_inference_steps (`int`):
103 | The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
104 | must be `None`.
105 | device (`str` or `torch.device`, *optional*):
106 | The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
107 | timesteps (`List[int]`, *optional*):
108 | Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
109 | `num_inference_steps` and `sigmas` must be `None`.
110 | sigmas (`List[float]`, *optional*):
111 | Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
112 | `num_inference_steps` and `timesteps` must be `None`.
113 |
114 | Returns:
115 | `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
116 | second element is the number of inference steps.
117 | """
118 | if timesteps is not None and sigmas is not None:
119 | raise ValueError(
120 | "Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values"
121 | )
122 | if timesteps is not None:
123 | accepts_timesteps = "timesteps" in set(
124 | inspect.signature(scheduler.set_timesteps).parameters.keys()
125 | )
126 | if not accepts_timesteps:
127 | raise ValueError(
128 | f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
129 | f" timestep schedules. Please check whether you are using the correct scheduler."
130 | )
131 | scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
132 | timesteps = scheduler.timesteps
133 | num_inference_steps = len(timesteps)
134 | elif sigmas is not None:
135 | accept_sigmas = "sigmas" in set(
136 | inspect.signature(scheduler.set_timesteps).parameters.keys()
137 | )
138 | if not accept_sigmas:
139 | raise ValueError(
140 | f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
141 | f" sigmas schedules. Please check whether you are using the correct scheduler."
142 | )
143 | scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
144 | timesteps = scheduler.timesteps
145 | num_inference_steps = len(timesteps)
146 | else:
147 | scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
148 | timesteps = scheduler.timesteps
149 | return timesteps, num_inference_steps
150 |
151 |
152 | class FluxConditionalPipeline(DiffusionPipeline, SD3LoraLoaderMixin):
153 | r"""
154 | The Flux pipeline for text-to-image generation.
155 |
156 | Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
157 |
158 | Args:
159 | transformer ([`FluxTransformer2DModel`]):
160 | Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
161 | scheduler ([`FlowMatchEulerDiscreteScheduler`]):
162 | A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
163 | vae ([`AutoencoderKL`]):
164 | Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
165 | text_encoder ([`CLIPTextModelWithProjection`]):
166 | [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
167 | specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
168 | with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
169 | as its dimension.
170 | text_encoder_2 ([`CLIPTextModelWithProjection`]):
171 | [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
172 | specifically the
173 | [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
174 | variant.
175 | tokenizer (`CLIPTokenizer`):
176 | Tokenizer of class
177 | [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
178 | tokenizer_2 (`CLIPTokenizer`):
179 | Second Tokenizer of class
180 | [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
181 | """
182 |
183 | model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
184 | _optional_components = []
185 | _callback_tensor_inputs = ["latents", "prompt_embeds"]
186 |
187 | def __init__(
188 | self,
189 | scheduler: FlowMatchEulerDiscreteScheduler,
190 | vae: AutoencoderKL,
191 | text_encoder: CLIPTextModel,
192 | tokenizer: CLIPTokenizer,
193 | text_encoder_2: T5EncoderModel,
194 | tokenizer_2: T5TokenizerFast,
195 | transformer: FluxTransformer2DConditionalModel,
196 | ):
197 | super().__init__()
198 |
199 | self.register_modules(
200 | vae=vae,
201 | text_encoder=text_encoder,
202 | text_encoder_2=text_encoder_2,
203 | tokenizer=tokenizer,
204 | tokenizer_2=tokenizer_2,
205 | transformer=transformer,
206 | scheduler=scheduler,
207 | )
208 | self.vae_scale_factor = (
209 | 2 ** (len(self.vae.config.block_out_channels))
210 | if hasattr(self, "vae") and self.vae is not None
211 | else 16
212 | )
213 | self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
214 | self.tokenizer_max_length = (
215 | self.tokenizer.model_max_length
216 | if hasattr(self, "tokenizer") and self.tokenizer is not None
217 | else 77
218 | )
219 | self.default_sample_size = 64
220 |
221 | def _get_t5_prompt_embeds(
222 | self,
223 | prompt: Union[str, List[str]] = None,
224 | num_images_per_prompt: int = 1,
225 | max_sequence_length: int = 512,
226 | device: Optional[torch.device] = None,
227 | dtype: Optional[torch.dtype] = None,
228 | ):
229 | device = device or self._execution_device
230 | dtype = dtype or self.text_encoder.dtype
231 |
232 | prompt = [prompt] if isinstance(prompt, str) else prompt
233 | batch_size = len(prompt)
234 |
235 | text_inputs = self.tokenizer_2(
236 | prompt,
237 | padding="max_length",
238 | max_length=max_sequence_length,
239 | truncation=True,
240 | return_length=False,
241 | return_overflowing_tokens=False,
242 | return_tensors="pt",
243 | )
244 | prompt_attention_mask = text_inputs.attention_mask
245 | text_input_ids = text_inputs.input_ids
246 | untruncated_ids = self.tokenizer_2(
247 | prompt, padding="longest", return_tensors="pt"
248 | ).input_ids
249 |
250 | if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
251 | text_input_ids, untruncated_ids
252 | ):
253 | removed_text = self.tokenizer_2.batch_decode(
254 | untruncated_ids[:, self.tokenizer_max_length - 1 : -1]
255 | )
256 | # logger.warning(
257 | # "The following part of your input was truncated because `max_sequence_length` is set to "
258 | # f" {max_sequence_length} tokens: {removed_text}"
259 | # )
260 |
261 | prompt_embeds = self.text_encoder_2(
262 | text_input_ids.to(device), output_hidden_states=False
263 | )[0]
264 |
265 | dtype = self.text_encoder_2.dtype
266 | prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
267 |
268 | _, seq_len, _ = prompt_embeds.shape
269 |
270 | # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
271 | prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
272 | prompt_embeds = prompt_embeds.view(
273 | batch_size * num_images_per_prompt, seq_len, -1
274 | )
275 |
276 | return prompt_embeds, prompt_attention_mask
277 |
278 | def _get_clip_prompt_embeds(
279 | self,
280 | prompt: Union[str, List[str]],
281 | num_images_per_prompt: int = 1,
282 | device: Optional[torch.device] = None,
283 | ):
284 | device = device or self._execution_device
285 |
286 | prompt = [prompt] if isinstance(prompt, str) else prompt
287 | batch_size = len(prompt)
288 |
289 | text_inputs = self.tokenizer(
290 | prompt,
291 | padding="max_length",
292 | max_length=self.tokenizer_max_length,
293 | truncation=True,
294 | return_overflowing_tokens=False,
295 | return_length=False,
296 | return_tensors="pt",
297 | )
298 |
299 | text_input_ids = text_inputs.input_ids
300 | untruncated_ids = self.tokenizer(
301 | prompt, padding="longest", return_tensors="pt"
302 | ).input_ids
303 | if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
304 | text_input_ids, untruncated_ids
305 | ):
306 | removed_text = self.tokenizer.batch_decode(
307 | untruncated_ids[:, self.tokenizer_max_length - 1 : -1]
308 | )
309 | logger.warning(
310 | "The following part of your input was truncated because CLIP can only handle sequences up to"
311 | f" {self.tokenizer_max_length} tokens: {removed_text}"
312 | )
313 | prompt_embeds = self.text_encoder(
314 | text_input_ids.to(device), output_hidden_states=False
315 | )
316 |
317 | # Use pooled output of CLIPTextModel
318 | prompt_embeds = prompt_embeds.pooler_output
319 | prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
320 |
321 | # duplicate text embeddings for each generation per prompt, using mps friendly method
322 | prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
323 | prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)
324 |
325 | return prompt_embeds
326 |
327 | def encode_prompt(
328 | self,
329 | prompt: Union[str, List[str]],
330 | prompt_2: Union[str, List[str]],
331 | device: Optional[torch.device] = None,
332 | num_images_per_prompt: int = 1,
333 | prompt_embeds: Optional[torch.FloatTensor] = None,
334 | pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
335 | max_sequence_length: int = 512,
336 | lora_scale: Optional[float] = None,
337 | ):
338 | r"""
339 |
340 | Args:
341 | prompt (`str` or `List[str]`, *optional*):
342 | prompt to be encoded
343 | prompt_2 (`str` or `List[str]`, *optional*):
344 | The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
345 | used in all text-encoders
346 | device: (`torch.device`):
347 | torch device
348 | num_images_per_prompt (`int`):
349 | number of images that should be generated per prompt
350 | prompt_embeds (`torch.FloatTensor`, *optional*):
351 | Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
352 | provided, text embeddings will be generated from `prompt` input argument.
353 | pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
354 | Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
355 | If not provided, pooled text embeddings will be generated from `prompt` input argument.
356 | clip_skip (`int`, *optional*):
357 | Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
358 | the output of the pre-final layer will be used for computing the prompt embeddings.
359 | lora_scale (`float`, *optional*):
360 | A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
361 | """
362 | device = device or self._execution_device
363 |
364 | # set lora scale so that monkey patched LoRA
365 | # function of text encoder can correctly access it
366 | if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
367 | self._lora_scale = lora_scale
368 |
369 | # dynamically adjust the LoRA scale
370 | if self.text_encoder is not None and USE_PEFT_BACKEND:
371 | scale_lora_layers(self.text_encoder, lora_scale)
372 | if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
373 | scale_lora_layers(self.text_encoder_2, lora_scale)
374 |
375 | prompt = [prompt] if isinstance(prompt, str) else prompt
376 | if prompt is not None:
377 | batch_size = len(prompt)
378 | else:
379 | batch_size = prompt_embeds.shape[0]
380 |
381 | prompt_attention_mask = None
382 | if prompt_embeds is None:
383 | prompt_2 = prompt_2 or prompt
384 | prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
385 |
386 | # We only use the pooled prompt output from the CLIPTextModel
387 | pooled_prompt_embeds = self._get_clip_prompt_embeds(
388 | prompt=prompt,
389 | device=device,
390 | num_images_per_prompt=num_images_per_prompt,
391 | )
392 | prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
393 | prompt=prompt_2,
394 | num_images_per_prompt=num_images_per_prompt,
395 | max_sequence_length=max_sequence_length,
396 | device=device
397 | )
398 |
399 | if self.text_encoder is not None:
400 | if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
401 | # Retrieve the original scale by scaling back the LoRA layers
402 | unscale_lora_layers(self.text_encoder, lora_scale)
403 |
404 | if self.text_encoder_2 is not None:
405 | if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
406 | # Retrieve the original scale by scaling back the LoRA layers
407 | unscale_lora_layers(self.text_encoder_2, lora_scale)
408 |
409 | text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(
410 | device=device, dtype=prompt_embeds.dtype
411 | )
412 |
413 | return prompt_embeds, pooled_prompt_embeds, text_ids, prompt_attention_mask
414 |
415 | def check_inputs(
416 | self,
417 | prompt,
418 | prompt_2,
419 | height,
420 | width,
421 | prompt_embeds=None,
422 | pooled_prompt_embeds=None,
423 | callback_on_step_end_tensor_inputs=None,
424 | max_sequence_length=None,
425 | image=None
426 | ):
427 | if height % 8 != 0 or width % 8 != 0:
428 | raise ValueError(
429 | f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
430 | )
431 |
432 | if callback_on_step_end_tensor_inputs is not None and not all(
433 | k in self._callback_tensor_inputs
434 | for k in callback_on_step_end_tensor_inputs
435 | ):
436 | raise ValueError(
437 | f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
438 | )
439 |
440 | if prompt is not None and prompt_embeds is not None:
441 | raise ValueError(
442 | f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
443 | " only forward one of the two."
444 | )
445 | elif prompt_2 is not None and prompt_embeds is not None:
446 | raise ValueError(
447 | f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
448 | " only forward one of the two."
449 | )
450 | elif prompt is None and prompt_embeds is None:
451 | raise ValueError(
452 | "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
453 | )
454 | elif prompt is not None and (
455 | not isinstance(prompt, str) and not isinstance(prompt, list)
456 | ):
457 | raise ValueError(
458 | f"`prompt` has to be of type `str` or `list` but is {type(prompt)}"
459 | )
460 | elif prompt_2 is not None and (
461 | not isinstance(prompt_2, str) and not isinstance(prompt_2, list)
462 | ):
463 | raise ValueError(
464 | f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}"
465 | )
466 |
467 | if prompt_embeds is not None and pooled_prompt_embeds is None:
468 | raise ValueError(
469 | "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
470 | )
471 |
472 | if max_sequence_length is not None and max_sequence_length > 512:
473 | raise ValueError(
474 | f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}"
475 | )
476 |
477 | @staticmethod
478 | def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
479 | latent_image_ids = torch.zeros(height // 2, width // 2, 3)
480 | latent_image_ids[..., 1] = (
481 | latent_image_ids[..., 1] + torch.arange(height // 2)[:, None]
482 | )
483 | latent_image_ids[..., 2] = (
484 | latent_image_ids[..., 2] + torch.arange(width // 2)[None, :]
485 | )
486 |
487 | latent_image_id_height, latent_image_id_width, latent_image_id_channels = (
488 | latent_image_ids.shape
489 | )
490 |
491 | latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
492 | latent_image_ids = latent_image_ids.reshape(
493 | batch_size,
494 | latent_image_id_height * latent_image_id_width,
495 | latent_image_id_channels,
496 | )
497 |
498 | return latent_image_ids.to(device=device, dtype=dtype)
499 |
500 | @staticmethod
501 | def _pack_latents(latents, batch_size, num_channels_latents, height, width):
502 | latents = latents.view(
503 | batch_size, num_channels_latents, height // 2, 2, width // 2, 2
504 | )
505 | latents = latents.permute(0, 2, 4, 1, 3, 5)
506 | latents = latents.reshape(
507 | batch_size, (height // 2) * (width // 2), num_channels_latents * 4
508 | )
509 |
510 | return latents
511 |
512 | @staticmethod
513 | def _unpack_latents(latents, height, width, vae_scale_factor):
514 | batch_size, num_patches, channels = latents.shape
515 |
516 | height = height // vae_scale_factor
517 | width = width // vae_scale_factor
518 |
519 | latents = latents.view(batch_size, height, width, channels // 4, 2, 2)
520 | latents = latents.permute(0, 3, 1, 4, 2, 5)
521 |
522 | latents = latents.reshape(
523 | batch_size, channels // (2 * 2), height * 2, width * 2
524 | )
525 |
526 | return latents
527 |
528 | def prepare_latents(
529 | self,
530 | batch_size,
531 | num_channels_latents,
532 | height,
533 | width,
534 | dtype,
535 | device,
536 | generator,
537 | latents=None,
538 | ):
539 | height = 2 * (int(height) // self.vae_scale_factor)
540 | width = 2 * (int(width) // self.vae_scale_factor)
541 |
542 | shape = (batch_size, num_channels_latents, height, width)
543 |
544 | if latents is not None:
545 | latent_image_ids = self._prepare_latent_image_ids(
546 | batch_size, height, width, device, dtype
547 | )
548 | return latents.to(device=device, dtype=dtype), latent_image_ids
549 |
550 | if isinstance(generator, list) and len(generator) != batch_size:
551 | raise ValueError(
552 | f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
553 | f" size of {batch_size}. Make sure the batch size matches the length of the generators."
554 | )
555 |
556 | latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
557 | # _pack_latents(latents, batch_size, num_channels_latents, height, width)
558 | latents = self._pack_latents(
559 | latents, batch_size, num_channels_latents, height, width
560 | )
561 |
562 | latent_image_ids = self._prepare_latent_image_ids(
563 | batch_size, height, width, device, dtype
564 | )
565 |
566 | return latents, latent_image_ids
567 |
568 | @property
569 | def guidance_scale(self):
570 | return self._guidance_scale
571 |
572 | @property
573 | def joint_attention_kwargs(self):
574 | return self._joint_attention_kwargs
575 |
576 | @property
577 | def num_timesteps(self):
578 | return self._num_timesteps
579 |
580 | @property
581 | def interrupt(self):
582 | return self._interrupt
583 |
584 | @torch.no_grad()
585 | @replace_example_docstring(EXAMPLE_DOC_STRING)
586 | def __call__(
587 | self,
588 | prompt: Union[str, List[str]] = None,
589 | prompt_mask: Optional[Union[torch.FloatTensor, List[torch.FloatTensor]]] = None,
590 | negative_mask: Optional[
591 | Union[torch.FloatTensor, List[torch.FloatTensor]]
592 | ] = None,
593 | prompt_2: Optional[Union[str, List[str]]] = None,
594 | height: Optional[int] = None,
595 | width: Optional[int] = None,
596 | num_inference_steps: int = 28,
597 | timesteps: List[int] = None,
598 | guidance_scale: float = 3.5,
599 | num_images_per_prompt: Optional[int] = 1,
600 | generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
601 | latents: Optional[torch.FloatTensor] = None,
602 | prompt_embeds: Optional[torch.FloatTensor] = None,
603 | pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
604 | output_type: Optional[str] = "pil",
605 | return_dict: bool = True,
606 | joint_attention_kwargs: Optional[Dict[str, Any]] = None,
607 | callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
608 | callback_on_step_end_tensor_inputs: List[str] = ["latents"],
609 | max_sequence_length: int = 512,
610 | guidance_scale_real_i: float = 1.0,
611 | guidance_scale_real_t: float = 1.0,
612 | negative_prompt: Union[str, List[str]] = "",
613 | negative_prompt_2: Union[str, List[str]] = "",
614 | negative_prompt_embeds: Optional[torch.FloatTensor] = None,
615 | negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
616 | no_cfg_until_timestep: int = 2,
617 | image: Optional[torch.FloatTensor] = None,
618 | cut_output = True
619 | ):
620 | r"""
621 | Function invoked when calling the pipeline for generation.
622 |
623 | Args:
624 | prompt (`str` or `List[str]`, *optional*):
625 | The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
626 | instead.
627 | prompt_mask (`str` or `List[str]`, *optional*):
628 | The prompt or prompts to be used as a mask for the image generation. If not defined, `prompt` is used
629 | instead.
630 | prompt_2 (`str` or `List[str]`, *optional*):
631 | The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
632 | will be used instead
633 | height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
634 | The height in pixels of the generated image. This is set to 1024 by default for the best results.
635 | width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
636 | The width in pixels of the generated image. This is set to 1024 by default for the best results.
637 | num_inference_steps (`int`, *optional*, defaults to 50):
638 | The number of denoising steps. More denoising steps usually lead to a higher quality image at the
639 | expense of slower inference.
640 | timesteps (`List[int]`, *optional*):
641 | Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
642 | in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
643 | passed will be used. Must be in descending order.
644 | guidance_scale (`float`, *optional*, defaults to 7.0):
645 | Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
646 | `guidance_scale` is defined as `w` of equation 2. of [Imagen
647 | Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
648 | 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
649 | usually at the expense of lower image quality.
650 | num_images_per_prompt (`int`, *optional*, defaults to 1):
651 | The number of images to generate per prompt.
652 | generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
653 | One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
654 | to make generation deterministic.
655 | latents (`torch.FloatTensor`, *optional*):
656 | Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
657 | generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
658 | tensor will ge generated by sampling using the supplied random `generator`.
659 | prompt_embeds (`torch.FloatTensor`, *optional*):
660 | Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
661 | provided, text embeddings will be generated from `prompt` input argument.
662 | pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
663 | Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
664 | If not provided, pooled text embeddings will be generated from `prompt` input argument.
665 | output_type (`str`, *optional*, defaults to `"pil"`):
666 | The output format of the generate image. Choose between
667 | [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
668 | return_dict (`bool`, *optional*, defaults to `True`):
669 | Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple.
670 | joint_attention_kwargs (`dict`, *optional*):
671 | A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
672 | `self.processor` in
673 | [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
674 | callback_on_step_end (`Callable`, *optional*):
675 | A function that calls at the end of each denoising steps during the inference. The function is called
676 | with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
677 | callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
678 | `callback_on_step_end_tensor_inputs`.
679 | callback_on_step_end_tensor_inputs (`List`, *optional*):
680 | The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
681 | will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
682 | `._callback_tensor_inputs` attribute of your pipeline class.
683 | max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
684 |
685 | Examples:
686 |
687 | Returns:
688 | [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict`
689 | is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated
690 | images.
691 | """
692 |
693 | height = height or self.default_sample_size * self.vae_scale_factor
694 | width = width or self.default_sample_size * self.vae_scale_factor
695 |
696 | # 1. Check inputs. Raise error if not correct
697 | self.check_inputs(
698 | prompt,
699 | prompt_2,
700 | height,
701 | width,
702 | prompt_embeds=prompt_embeds,
703 | pooled_prompt_embeds=pooled_prompt_embeds,
704 | callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
705 | max_sequence_length=max_sequence_length,
706 | )
707 |
708 | self._guidance_scale = guidance_scale
709 | self._guidance_scale_real_i = guidance_scale_real_i
710 | self._guidance_scale_real_t = guidance_scale_real_t
711 | self._joint_attention_kwargs = joint_attention_kwargs
712 | self._interrupt = False
713 |
714 | # 2. Define call parameters
715 | if prompt is not None and isinstance(prompt, str):
716 | batch_size = 1
717 | elif prompt is not None and isinstance(prompt, list):
718 | batch_size = len(prompt)
719 | else:
720 | batch_size = prompt_embeds.shape[0]
721 |
722 | device = self._execution_device
723 |
724 | # prompt = enhance_prompt(image, prompt)
725 | # if gemini_prompt:
726 | # while True:
727 | # try:
728 | # prompt = enhance_prompt(image, prompt)
729 | # break # Exit the loop if the function succeeds
730 | # except Exception as e:
731 | # print(f"An error occurred: {e}")
732 |
733 | lora_scale = (
734 | self.joint_attention_kwargs.get("scale", None)
735 | if self.joint_attention_kwargs is not None
736 | else None
737 | )
738 | (
739 | prompt_embeds,
740 | pooled_prompt_embeds,
741 | text_ids,
742 | _,
743 | ) = self.encode_prompt(
744 | prompt=prompt,
745 | prompt_2=prompt_2,
746 | prompt_embeds=prompt_embeds,
747 | pooled_prompt_embeds=pooled_prompt_embeds,
748 | device=device,
749 | num_images_per_prompt=num_images_per_prompt,
750 | max_sequence_length=max_sequence_length,
751 | lora_scale=lora_scale,
752 | )
753 |
754 | if negative_prompt_2 == "" and negative_prompt != "":
755 | negative_prompt_2 = negative_prompt
756 |
757 | negative_text_ids = text_ids
758 | if guidance_scale_real_i > 1.0 and (
759 | negative_prompt_embeds is None or negative_pooled_prompt_embeds is None
760 | ):
761 | (
762 | negative_prompt_embeds,
763 | negative_pooled_prompt_embeds,
764 | negative_text_ids,
765 | _,
766 | ) = self.encode_prompt(
767 | prompt=negative_prompt,
768 | prompt_2=negative_prompt_2,
769 | prompt_embeds=None,
770 | pooled_prompt_embeds=None,
771 | device=device,
772 | num_images_per_prompt=num_images_per_prompt,
773 | max_sequence_length=max_sequence_length,
774 | lora_scale=lora_scale,
775 | )
776 |
777 | # 3. Preprocess image
778 | image = self.image_processor.preprocess(image)
779 | # image = image[..., :512]
780 | image = torch.nn.functional.interpolate(image, size=(height, width // 2))
781 | black_image = torch.full((1, 3, height, width // 2), -1.0)
782 | image = torch.cat([image, black_image], dim=3)
783 | latents_cond = self.vae.encode(image.to(dtype=self.vae.dtype).to(device)).latent_dist.sample()
784 | latents_cond = (
785 | latents_cond - self.vae.config.shift_factor
786 | ) * self.vae.config.scaling_factor
787 | # from customization.utils import mask_random_quadrants
788 | # latent_cond = mask_random_quadrants(latent_cond)
789 |
790 | # 4. Prepare latent variables
791 | num_channels_latents = self.transformer.config.in_channels // 4
792 | latents, latent_image_ids = self.prepare_latents(
793 | batch_size * num_images_per_prompt,
794 | num_channels_latents,
795 | height,
796 | width,
797 | prompt_embeds.dtype,
798 | device,
799 | generator,
800 | latents,
801 | )
802 | # _pack_latents(latents, batch_size, num_channels_latents, height, width)
803 | latents_cond = self._pack_latents(
804 | latents_cond, batch_size, num_channels_latents, 2 * (int(height) // self.vae_scale_factor), 2 * (int(width) // self.vae_scale_factor)
805 | )
806 |
807 | # 5. Prepare timesteps
808 | sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
809 | image_seq_len = latents.shape[1]
810 | mu = calculate_shift(
811 | image_seq_len,
812 | self.scheduler.config.base_image_seq_len,
813 | self.scheduler.config.max_image_seq_len,
814 | self.scheduler.config.base_shift,
815 | self.scheduler.config.max_shift,
816 | )
817 | timesteps, num_inference_steps = retrieve_timesteps(
818 | self.scheduler,
819 | num_inference_steps,
820 | device,
821 | timesteps,
822 | sigmas,
823 | mu=mu,
824 | )
825 | num_warmup_steps = max(
826 | len(timesteps) - num_inference_steps * self.scheduler.order, 0
827 | )
828 | self._num_timesteps = len(timesteps)
829 |
830 | # 6. Denoising loop
831 | with self.progress_bar(total=num_inference_steps) as progress_bar:
832 | for i, t in enumerate(timesteps):
833 | if self.interrupt:
834 | continue
835 |
836 | # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
837 | timestep = t.expand(latents.shape[0]).to(latents.dtype)
838 |
839 | # handle guidance
840 | if self.transformer.config.guidance_embeds:
841 | guidance = torch.tensor(
842 | [guidance_scale], device=device
843 | )
844 | guidance = guidance.expand(latents.shape[0])
845 | else:
846 | guidance = None
847 |
848 | extra_transformer_args = {}
849 | if prompt_mask is not None:
850 | extra_transformer_args["attention_mask"] = prompt_mask.to(
851 | device=self.transformer.device
852 | )
853 |
854 | noise_pred = self.transformer(
855 | hidden_states=latents,
856 | # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
857 | timestep=timestep / 1000,
858 | guidance=guidance,
859 | pooled_projections=pooled_prompt_embeds,
860 | encoder_hidden_states=prompt_embeds,
861 | txt_ids=text_ids,
862 | img_ids=latent_image_ids,
863 | joint_attention_kwargs=self.joint_attention_kwargs,
864 | return_dict=False,
865 | condition_hidden_states=latents_cond,
866 | **extra_transformer_args,
867 | )[0]
868 |
869 | # TODO optionally use batch prediction to speed this up.
870 | if guidance_scale_real_i > 1.0 and i >= no_cfg_until_timestep:
871 | noise_pred_uncond = self.transformer(
872 | hidden_states=latents,
873 | # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
874 | timestep=timestep / 1000,
875 | guidance=guidance,
876 | pooled_projections=negative_pooled_prompt_embeds,
877 | encoder_hidden_states=negative_prompt_embeds,
878 | txt_ids=negative_text_ids,
879 | img_ids=latent_image_ids,
880 | joint_attention_kwargs=self.joint_attention_kwargs,
881 | return_dict=False,
882 | condition_hidden_states=torch.zeros_like(latents_cond),
883 | )[0]
884 | noise_pred_uncond_t = self.transformer(
885 | hidden_states=latents,
886 | # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
887 | timestep=timestep / 1000,
888 | guidance=guidance,
889 | pooled_projections=negative_pooled_prompt_embeds,
890 | encoder_hidden_states=negative_prompt_embeds,
891 | txt_ids=negative_text_ids,
892 | img_ids=latent_image_ids,
893 | joint_attention_kwargs=self.joint_attention_kwargs,
894 | return_dict=False,
895 | condition_hidden_states=latents_cond,
896 | )[0]
897 |
898 | # noise_pred = noise_pred_uncond + guidance_scale_real * (
899 | # noise_pred - noise_pred_uncond
900 | # )
901 | noise_pred = (
902 | noise_pred_uncond
903 | + guidance_scale_real_i
904 | * (noise_pred_uncond_t - noise_pred_uncond)
905 | + guidance_scale_real_t * (noise_pred - noise_pred_uncond_t)
906 | )
907 |
908 | # compute the previous noisy sample x_t -> x_t-1
909 | latents_dtype = latents.dtype
910 | latents = self.scheduler.step(
911 | noise_pred, t, latents, return_dict=False
912 | )[0]
913 |
914 | if latents.dtype != latents_dtype:
915 | if torch.backends.mps.is_available():
916 | # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
917 | latents = latents.to(latents_dtype)
918 |
919 | if callback_on_step_end is not None:
920 | callback_kwargs = {}
921 | for k in callback_on_step_end_tensor_inputs:
922 | callback_kwargs[k] = locals()[k]
923 | callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
924 |
925 | latents = callback_outputs.pop("latents", latents)
926 | prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
927 |
928 | # call the callback, if provided
929 | if i == len(timesteps) - 1 or (
930 | (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
931 | ):
932 | progress_bar.update()
933 |
934 | if XLA_AVAILABLE:
935 | xm.mark_step()
936 |
937 | reference_image = None
938 |
939 | if output_type == "latent":
940 | image = latents
941 |
942 | else:
943 | latents = self._unpack_latents(
944 | latents, height, width, self.vae_scale_factor
945 | )
946 | latents = (
947 | latents / self.vae.config.scaling_factor
948 | ) + self.vae.config.shift_factor
949 |
950 | image = self.vae.decode(
951 | latents,
952 | return_dict=False,
953 | )[0]
954 |
955 | # Debug the image shape before splitting
956 | print(f"Image shape before splitting: {image.shape}")
957 |
958 | if cut_output:
959 | # Store the reference image (left side)
960 | reference_image = image[...,:width//2].clone() # Use clone to ensure we have a separate copy
961 | # Store the output image (right side)
962 | image = image[..., width//2:]
963 |
964 | # Debug the shapes after splitting
965 | print(f"Reference image shape after splitting: {reference_image.shape}")
966 | print(f"Output image shape after splitting: {image.shape}")
967 |
968 | # Post-process the images
969 | image = self.image_processor.postprocess(image, output_type=output_type)
970 | if reference_image is not None:
971 | reference_image = self.image_processor.postprocess(reference_image, output_type=output_type)
972 | print(f"Reference image type after postprocessing: {type(reference_image)}")
973 | if isinstance(reference_image, list):
974 | print(f"Reference image list length: {len(reference_image)}")
975 | if len(reference_image) > 0:
976 | print(f"First reference image type: {type(reference_image[0])}")
977 | print(f"First reference image size: {reference_image[0].size if hasattr(reference_image[0], 'size') else 'unknown'}")
978 |
979 | # Offload all models
980 | self.maybe_free_model_hooks()
981 |
982 | if not return_dict:
983 | return (image, reference_image)
984 |
985 | return FluxPipelineOutput(images=(image, reference_image))
986 |
987 |
988 | from dataclasses import dataclass
989 | from typing import List, Union
990 | import PIL.Image
991 | from diffusers.utils import BaseOutput
992 |
993 |
994 | @dataclass
995 | class FluxPipelineOutput(BaseOutput):
996 | """
997 | Output class for Stable Diffusion pipelines.
998 |
999 | Args:
1000 | images (`List[PIL.Image.Image]` or `np.ndarray` or tuple)
1001 | List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
1002 | num_channels)` or tuple of (output_image, reference_image). PIL images or numpy array present the denoised
1003 | images of the diffusion pipeline.
1004 | """
1005 |
1006 | images: Union[List[PIL.Image.Image], np.ndarray, tuple]
1007 |
1008 | def __post_init__(self):
1009 | # Ensure images is always a tuple of (output_image, reference_image)
1010 | if not isinstance(self.images, tuple):
1011 | self.images = (self.images, None)
1012 |
--------------------------------------------------------------------------------
/core/recaption.py:
--------------------------------------------------------------------------------
1 | from google import genai
2 |
3 |
4 | def enhance_prompt(image, prompt,api_key):
5 | # input_caption_prompt = "Please provide a prompt for the image for Diffusion Model text-to-image generative model training, i.e. for FLUX or StableDiffusion 3. The prompt should be a detailed description of the image, including the character/asset/item, the environment, the pose, the lighting, the camera view, etc. The prompt should be detailed enough to generate the image. The prompt should be as short and precise as possible, in one-line format, and does not exceed 77 tokens."
6 | input_caption_prompt = (
7 | "Please provide a prompt for a Diffusion Model text-to-image generative model for the image I will give you. "
8 | "The prompt should be a detailed description of the image, especially the main subject (i.e. the main character/asset/item), the environment, the pose, the lighting, the camera view, the style etc."
9 | "The prompt should be detailed enough to generate the target image. "
10 | # "Identify key elements and that remain consistent from the source image, and highlight differences in the target image. "
11 | # "The prompt should be short, precise, one-line, similar to LAION dataset style, and not exceed 77 tokens. "
12 | # "The prompt should be detailed enough to generate the target image."
13 | "The prompt should be short and precise, in one-line format, and does not exceed 77 tokens."
14 | "The prompt should be individually coherent as a description of the image."
15 | )
16 |
17 | # Choose a Gemini model.
18 | caption_model = genai.Client(api_key=api_key)
19 | input_image_prompt = caption_model.models.generate_content(
20 | model='gemini-2.0-flash-001', contents=[input_caption_prompt, image]).text
21 | input_image_prompt = input_image_prompt.replace('\r', '').replace('\n', '')
22 |
23 | enhance_instruction = "Enhance this input text prompt: '"
24 | enhance_instruction += prompt
25 | enhance_instruction += "'. Please extract other details, especially description of the main subject from the following reference prompt: '"
26 | enhance_instruction += input_image_prompt
27 | enhance_instruction += "'. Please keep the details that are mentioned in the input prompt, and enhance the rest. "
28 | enhance_instruction += "Response with only the enhanced prompt. "
29 | enhance_instruction += "The enhanced prompt should be short and precise, in one-line format, and does not exceed 77 tokens."
30 | enhanced_prompt = caption_model.models.generate_content(
31 | model='gemini-2.0-flash-001', contents=[enhance_instruction]).text.replace('\r', '').replace('\n', '')
32 | print("input_image_prompt: ", input_image_prompt)
33 | print("prompt: ", prompt)
34 | print("enhanced_prompt: ", enhanced_prompt)
35 | return enhanced_prompt
36 |
37 |
38 |
--------------------------------------------------------------------------------
/core/transformer.py:
--------------------------------------------------------------------------------
1 | # Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
2 | #
3 | # Licensed under the Apache License, Version 2.0 (the "License");
4 | # you may not use this file except in compliance with the License.
5 | # You may obtain a copy of the License at
6 | #
7 | # http://www.apache.org/licenses/LICENSE-2.0
8 | #
9 | # Unless required by applicable law or agreed to in writing, software
10 | # distributed under the License is distributed on an "AS IS" BASIS,
11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 | # See the License for the specific language governing permissions and
13 | # limitations under the License.
14 |
15 |
16 | from typing import Any, Dict, Optional, Tuple, Union
17 |
18 | import numpy as np
19 | import torch
20 | import torch.nn as nn
21 | import torch.nn.functional as F
22 |
23 | from diffusers.configuration_utils import ConfigMixin, register_to_config
24 | from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
25 | from diffusers.models.attention import FeedForward
26 | from diffusers.models.attention_processor import (
27 | Attention,
28 | AttentionProcessor,
29 | FluxAttnProcessor2_0,
30 | FusedFluxAttnProcessor2_0,
31 | )
32 | from diffusers.models.modeling_utils import ModelMixin
33 | from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
34 | from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
35 | from diffusers.utils.torch_utils import maybe_allow_in_graph
36 | from diffusers.models.embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
37 | from diffusers.models.modeling_outputs import Transformer2DModelOutput
38 |
39 |
40 | logger = logging.get_logger(__name__) # pylint: disable=invalid-name
41 |
42 |
43 | @maybe_allow_in_graph
44 | class FluxSingleTransformerBlock(nn.Module):
45 | r"""
46 | A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
47 |
48 | Reference: https://arxiv.org/abs/2403.03206
49 |
50 | Parameters:
51 | dim (`int`): The number of channels in the input and output.
52 | num_attention_heads (`int`): The number of heads to use for multi-head attention.
53 | attention_head_dim (`int`): The number of channels in each head.
54 | context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
55 | processing of `context` conditions.
56 | """
57 |
58 | def __init__(self, dim, num_attention_heads, attention_head_dim, mlp_ratio=4.0):
59 | super().__init__()
60 | self.mlp_hidden_dim = int(dim * mlp_ratio)
61 |
62 | self.norm = AdaLayerNormZeroSingle(dim)
63 | self.proj_mlp = nn.Linear(dim, self.mlp_hidden_dim)
64 | self.act_mlp = nn.GELU(approximate="tanh")
65 | self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
66 |
67 | processor = FluxAttnProcessor2_0()
68 | self.attn = Attention(
69 | query_dim=dim,
70 | cross_attention_dim=None,
71 | dim_head=attention_head_dim,
72 | heads=num_attention_heads,
73 | out_dim=dim,
74 | bias=True,
75 | processor=processor,
76 | qk_norm="rms_norm",
77 | eps=1e-6,
78 | pre_only=True,
79 | )
80 |
81 | def forward(
82 | self,
83 | hidden_states: torch.FloatTensor,
84 | temb: torch.FloatTensor,
85 | image_rotary_emb=None,
86 | ):
87 | residual = hidden_states
88 | norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
89 | mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
90 |
91 | attn_output = self.attn(
92 | hidden_states=norm_hidden_states,
93 | image_rotary_emb=image_rotary_emb,
94 | )
95 |
96 | hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
97 | gate = gate.unsqueeze(1)
98 | hidden_states = gate * self.proj_out(hidden_states)
99 | hidden_states = residual + hidden_states
100 | if hidden_states.dtype == torch.float16:
101 | hidden_states = hidden_states.clip(-65504, 65504)
102 |
103 | return hidden_states
104 |
105 |
106 | @maybe_allow_in_graph
107 | class FluxTransformerBlock(nn.Module):
108 | r"""
109 | A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
110 |
111 | Reference: https://arxiv.org/abs/2403.03206
112 |
113 | Parameters:
114 | dim (`int`): The number of channels in the input and output.
115 | num_attention_heads (`int`): The number of heads to use for multi-head attention.
116 | attention_head_dim (`int`): The number of channels in each head.
117 | context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
118 | processing of `context` conditions.
119 | """
120 |
121 | def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm="rms_norm", eps=1e-6):
122 | super().__init__()
123 |
124 | self.norm1 = AdaLayerNormZero(dim)
125 |
126 | self.norm1_context = AdaLayerNormZero(dim)
127 |
128 | if hasattr(F, "scaled_dot_product_attention"):
129 | processor = FluxAttnProcessor2_0()
130 | else:
131 | raise ValueError(
132 | "The current PyTorch version does not support the `scaled_dot_product_attention` function."
133 | )
134 | self.attn = Attention(
135 | query_dim=dim,
136 | cross_attention_dim=None,
137 | added_kv_proj_dim=dim,
138 | dim_head=attention_head_dim,
139 | heads=num_attention_heads,
140 | out_dim=dim,
141 | context_pre_only=False,
142 | bias=True,
143 | processor=processor,
144 | qk_norm=qk_norm,
145 | eps=eps,
146 | )
147 |
148 | self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
149 | self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
150 |
151 | self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
152 | self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
153 |
154 | # let chunk size default to None
155 | self._chunk_size = None
156 | self._chunk_dim = 0
157 |
158 | def forward(
159 | self,
160 | hidden_states: torch.FloatTensor,
161 | encoder_hidden_states: torch.FloatTensor,
162 | temb: torch.FloatTensor,
163 | image_rotary_emb=None,
164 | ):
165 | norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
166 |
167 | norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
168 | encoder_hidden_states, emb=temb
169 | )
170 |
171 | # Attention.
172 | attn_output, context_attn_output = self.attn(
173 | hidden_states=norm_hidden_states,
174 | encoder_hidden_states=norm_encoder_hidden_states,
175 | image_rotary_emb=image_rotary_emb,
176 | )
177 |
178 | # Process attention outputs for the `hidden_states`.
179 | attn_output = gate_msa.unsqueeze(1) * attn_output
180 | hidden_states = hidden_states + attn_output
181 |
182 | norm_hidden_states = self.norm2(hidden_states)
183 | norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
184 |
185 | ff_output = self.ff(norm_hidden_states)
186 | ff_output = gate_mlp.unsqueeze(1) * ff_output
187 |
188 | hidden_states = hidden_states + ff_output
189 |
190 | # Process attention outputs for the `encoder_hidden_states`.
191 |
192 | context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
193 | encoder_hidden_states = encoder_hidden_states + context_attn_output
194 |
195 | norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
196 | norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
197 |
198 | context_ff_output = self.ff_context(norm_encoder_hidden_states)
199 | encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
200 | if encoder_hidden_states.dtype == torch.float16:
201 | encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
202 |
203 | return encoder_hidden_states, hidden_states
204 |
205 |
206 | class FluxTransformer2DConditionalModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
207 | """
208 | The Transformer model introduced in Flux.
209 |
210 | Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
211 |
212 | Parameters:
213 | patch_size (`int`): Patch size to turn the input data into small patches.
214 | in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
215 | num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
216 | num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
217 | attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
218 | num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
219 | joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
220 | pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
221 | guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
222 | """
223 |
224 | _supports_gradient_checkpointing = True
225 | _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
226 |
227 | @register_to_config
228 | def __init__(
229 | self,
230 | patch_size: int = 1,
231 | in_channels: int = 64,
232 | num_layers: int = 19,
233 | num_single_layers: int = 38,
234 | attention_head_dim: int = 128,
235 | num_attention_heads: int = 24,
236 | joint_attention_dim: int = 4096,
237 | pooled_projection_dim: int = 768,
238 | guidance_embeds: bool = False,
239 | axes_dims_rope: Tuple[int] = (16, 56, 56),
240 | ):
241 | super().__init__()
242 | self.out_channels = in_channels
243 | self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
244 |
245 | self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
246 |
247 | text_time_guidance_cls = (
248 | CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
249 | )
250 | self.time_text_embed = text_time_guidance_cls(
251 | embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_dim
252 | )
253 |
254 | self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
255 | self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)
256 | self.c_embedder = zero_module(torch.nn.Linear(self.config.in_channels, self.inner_dim))
257 |
258 | self.transformer_blocks = nn.ModuleList(
259 | [
260 | FluxTransformerBlock(
261 | dim=self.inner_dim,
262 | num_attention_heads=self.config.num_attention_heads,
263 | attention_head_dim=self.config.attention_head_dim,
264 | )
265 | for i in range(self.config.num_layers)
266 | ]
267 | )
268 |
269 | self.single_transformer_blocks = nn.ModuleList(
270 | [
271 | FluxSingleTransformerBlock(
272 | dim=self.inner_dim,
273 | num_attention_heads=self.config.num_attention_heads,
274 | attention_head_dim=self.config.attention_head_dim,
275 | )
276 | for i in range(self.config.num_single_layers)
277 | ]
278 | )
279 |
280 | self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
281 | self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
282 |
283 | self.gradient_checkpointing = False
284 |
285 | @property
286 | # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
287 | def attn_processors(self) -> Dict[str, AttentionProcessor]:
288 | r"""
289 | Returns:
290 | `dict` of attention processors: A dictionary containing all attention processors used in the model with
291 | indexed by its weight name.
292 | """
293 | # set recursively
294 | processors = {}
295 |
296 | def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
297 | if hasattr(module, "get_processor"):
298 | processors[f"{name}.processor"] = module.get_processor()
299 |
300 | for sub_name, child in module.named_children():
301 | fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
302 |
303 | return processors
304 |
305 | for name, module in self.named_children():
306 | fn_recursive_add_processors(name, module, processors)
307 |
308 | return processors
309 |
310 | # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
311 | def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
312 | r"""
313 | Sets the attention processor to use to compute attention.
314 |
315 | Parameters:
316 | processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
317 | The instantiated processor class or a dictionary of processor classes that will be set as the processor
318 | for **all** `Attention` layers.
319 |
320 | If `processor` is a dict, the key needs to define the path to the corresponding cross attention
321 | processor. This is strongly recommended when setting trainable attention processors.
322 |
323 | """
324 | count = len(self.attn_processors.keys())
325 |
326 | if isinstance(processor, dict) and len(processor) != count:
327 | raise ValueError(
328 | f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
329 | f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
330 | )
331 |
332 | def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
333 | if hasattr(module, "set_processor"):
334 | if not isinstance(processor, dict):
335 | module.set_processor(processor)
336 | else:
337 | module.set_processor(processor.pop(f"{name}.processor"))
338 |
339 | for sub_name, child in module.named_children():
340 | fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
341 |
342 | for name, module in self.named_children():
343 | fn_recursive_attn_processor(name, module, processor)
344 |
345 | # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
346 | def fuse_qkv_projections(self):
347 | """
348 | Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
349 | are fused. For cross-attention modules, key and value projection matrices are fused.
350 |
351 |
352 |
353 | This API is 🧪 experimental.
354 |
355 |
356 | """
357 | self.original_attn_processors = None
358 |
359 | for _, attn_processor in self.attn_processors.items():
360 | if "Added" in str(attn_processor.__class__.__name__):
361 | raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
362 |
363 | self.original_attn_processors = self.attn_processors
364 |
365 | for module in self.modules():
366 | if isinstance(module, Attention):
367 | module.fuse_projections(fuse=True)
368 |
369 | self.set_attn_processor(FusedFluxAttnProcessor2_0())
370 |
371 | # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
372 | def unfuse_qkv_projections(self):
373 | """Disables the fused QKV projection if enabled.
374 |
375 |
376 |
377 | This API is 🧪 experimental.
378 |
379 |
380 |
381 | """
382 | if self.original_attn_processors is not None:
383 | self.set_attn_processor(self.original_attn_processors)
384 |
385 | def _set_gradient_checkpointing(self, module, value=False):
386 | if hasattr(module, "gradient_checkpointing"):
387 | module.gradient_checkpointing = value
388 |
389 | def forward(
390 | self,
391 | hidden_states: torch.Tensor,
392 | encoder_hidden_states: torch.Tensor = None,
393 | pooled_projections: torch.Tensor = None,
394 | timestep: torch.LongTensor = None,
395 | img_ids: torch.Tensor = None,
396 | txt_ids: torch.Tensor = None,
397 | guidance: torch.Tensor = None,
398 | joint_attention_kwargs: Optional[Dict[str, Any]] = None,
399 | controlnet_block_samples=None,
400 | controlnet_single_block_samples=None,
401 | condition_hidden_states: torch.Tensor = None,
402 | return_dict: bool = True,
403 | ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
404 | """
405 | The [`FluxTransformer2DModel`] forward method.
406 |
407 | Args:
408 | hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
409 | Input `hidden_states`.
410 | encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
411 | Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
412 | pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
413 | from the embeddings of input conditions.
414 | timestep ( `torch.LongTensor`):
415 | Used to indicate denoising step.
416 | block_controlnet_hidden_states: (`list` of `torch.Tensor`):
417 | A list of tensors that if specified are added to the residuals of transformer blocks.
418 | joint_attention_kwargs (`dict`, *optional*):
419 | A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
420 | `self.processor` in
421 | [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
422 | return_dict (`bool`, *optional*, defaults to `True`):
423 | Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
424 | tuple.
425 |
426 | Returns:
427 | If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
428 | `tuple` where the first element is the sample tensor.
429 | """
430 | if joint_attention_kwargs is not None:
431 | joint_attention_kwargs = joint_attention_kwargs.copy()
432 | lora_scale = joint_attention_kwargs.pop("scale", 1.0)
433 | else:
434 | lora_scale = 1.0
435 |
436 | if USE_PEFT_BACKEND:
437 | # weight the lora layers by setting `lora_scale` for each PEFT layer
438 | scale_lora_layers(self, lora_scale)
439 | else:
440 | if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
441 | logger.warning(
442 | "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
443 | )
444 | hidden_states = self.x_embedder(hidden_states) + self.c_embedder(condition_hidden_states)
445 |
446 | timestep = timestep.to(hidden_states.dtype) * 1000
447 | if guidance is not None:
448 | guidance = guidance.to(hidden_states.dtype) * 1000
449 | else:
450 | guidance = None
451 | temb = (
452 | self.time_text_embed(timestep, pooled_projections)
453 | if guidance is None
454 | else self.time_text_embed(timestep, guidance, pooled_projections)
455 | )
456 | encoder_hidden_states = self.context_embedder(encoder_hidden_states)
457 |
458 | if txt_ids.ndim == 3:
459 | # logger.warning(
460 | # "Passing `txt_ids` 3d torch.Tensor is deprecated."
461 | # "Please remove the batch dimension and pass it as a 2d torch Tensor"
462 | # )
463 | txt_ids = txt_ids[0]
464 | if img_ids.ndim == 3:
465 | # logger.warning(
466 | # "Passing `img_ids` 3d torch.Tensor is deprecated."
467 | # "Please remove the batch dimension and pass it as a 2d torch Tensor"
468 | # )
469 | img_ids = img_ids[0]
470 | ids = torch.cat((txt_ids, img_ids), dim=0)
471 | image_rotary_emb = self.pos_embed(ids)
472 |
473 | for index_block, block in enumerate(self.transformer_blocks):
474 | if self.training and self.gradient_checkpointing:
475 |
476 | def create_custom_forward(module, return_dict=None):
477 | def custom_forward(*inputs):
478 | if return_dict is not None:
479 | return module(*inputs, return_dict=return_dict)
480 | else:
481 | return module(*inputs)
482 |
483 | return custom_forward
484 |
485 | ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
486 | encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
487 | create_custom_forward(block),
488 | hidden_states,
489 | encoder_hidden_states,
490 | temb,
491 | image_rotary_emb,
492 | **ckpt_kwargs,
493 | )
494 |
495 | else:
496 | encoder_hidden_states, hidden_states = block(
497 | hidden_states=hidden_states,
498 | encoder_hidden_states=encoder_hidden_states,
499 | temb=temb,
500 | image_rotary_emb=image_rotary_emb,
501 | )
502 |
503 | # controlnet residual
504 | if controlnet_block_samples is not None:
505 | interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
506 | interval_control = int(np.ceil(interval_control))
507 | hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
508 |
509 | hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
510 |
511 | for index_block, block in enumerate(self.single_transformer_blocks):
512 | if self.training and self.gradient_checkpointing:
513 |
514 | def create_custom_forward(module, return_dict=None):
515 | def custom_forward(*inputs):
516 | if return_dict is not None:
517 | return module(*inputs, return_dict=return_dict)
518 | else:
519 | return module(*inputs)
520 |
521 | return custom_forward
522 |
523 | ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
524 | hidden_states = torch.utils.checkpoint.checkpoint(
525 | create_custom_forward(block),
526 | hidden_states,
527 | temb,
528 | image_rotary_emb,
529 | **ckpt_kwargs,
530 | )
531 |
532 | else:
533 | hidden_states = block(
534 | hidden_states=hidden_states,
535 | temb=temb,
536 | image_rotary_emb=image_rotary_emb,
537 | )
538 |
539 | # controlnet residual
540 | if controlnet_single_block_samples is not None:
541 | interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
542 | interval_control = int(np.ceil(interval_control))
543 | hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
544 | hidden_states[:, encoder_hidden_states.shape[1] :, ...]
545 | + controlnet_single_block_samples[index_block // interval_control]
546 | )
547 |
548 | hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
549 |
550 | hidden_states = self.norm_out(hidden_states, temb)
551 | output = self.proj_out(hidden_states)
552 |
553 | if USE_PEFT_BACKEND:
554 | # remove `lora_scale` from each PEFT layer
555 | unscale_lora_layers(self, lora_scale)
556 |
557 | if not return_dict:
558 | return (output,)
559 |
560 | return Transformer2DModelOutput(sample=output)
561 |
562 | def zero_module(module):
563 | for p in module.parameters():
564 | torch.nn.init.zeros_(p)
565 | return module
--------------------------------------------------------------------------------
/dsd_imports.py:
--------------------------------------------------------------------------------
1 | import os
2 | import sys
3 | import importlib
4 |
5 |
6 |
7 | from .core.pipeline import FluxConditionalPipeline
8 | from .core.transformer import FluxTransformer2DConditionalModel
9 | from .core.recaption import enhance_prompt
10 | IMPORTS_AVAILABLE = True
11 | print("Successfully imported DSD components from the node's core module.")
12 |
--------------------------------------------------------------------------------
/dsd_nodes.py:
--------------------------------------------------------------------------------
1 | import os
2 | import torch
3 | import folder_paths
4 | from PIL import Image
5 | import numpy as np
6 | import shutil
7 | import requests
8 | import json
9 | from pathlib import Path
10 | from tqdm import tqdm
11 | from huggingface_hub import hf_hub_download, snapshot_download
12 |
13 | from .utils import get_model_path, get_lora_path, comfy_to_pil, pil_to_comfy, resize_and_center_crop, center_crop, pad_resize, fit_resize
14 | from .dsd_imports import FluxConditionalPipeline, FluxTransformer2DConditionalModel, enhance_prompt, IMPORTS_AVAILABLE
15 | from comfy.utils import ProgressBar
16 | try:
17 | from google import genai
18 | GEMINI_AVAILABLE = True
19 | except ImportError:
20 | GEMINI_AVAILABLE = False
21 | print("Google Gemini API not available. Install with: pip install google-genai")
22 |
23 | # Add support paths
24 | custom_nodes_dir = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
25 |
26 | # Add ComfyUI models path
27 | comfyui_models_path = os.path.join(os.path.dirname(custom_nodes_dir), "models")
28 | os.makedirs(comfyui_models_path, exist_ok=True)
29 | dsd_model_path = os.path.join(comfyui_models_path, "dsd_model")
30 | os.makedirs(dsd_model_path, exist_ok=True)
31 |
32 | # Register the dsd_model path with ComfyUI
33 | folder_paths.add_model_folder_path("dsd_models", dsd_model_path)
34 |
35 | class DSDModelLoader:
36 | """Loads the DSD (Diffusion Self-Distillation) model"""
37 |
38 | @classmethod
39 | def INPUT_TYPES(cls):
40 | return {
41 | "required": {
42 | "model_path": ("STRING", {"default": ""}),
43 | "lora_path": ("STRING", {"default": ""}),
44 | "device": (["cuda", "cpu"], {"default": "cuda"}),
45 | "dtype": (["bfloat16", "float16", "float32"], {"default": "bfloat16"}),
46 | "low_cpu_mem_usage": ("BOOLEAN", {"default": True, "tooltip": "Reduces CPU memory usage during model loading. Recommended for faster loading."}),
47 | "model_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "Offloads state dict to reduce memory usage during loading. May slow down inference speed."}),
48 | "sequential_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "Enables sequential CPU offloading. Only use if low on VRAM. Significantly impacts inference speed."})
49 | }
50 | }
51 |
52 | RETURN_TYPES = ("DSD_MODEL",)
53 | RETURN_NAMES = ("dsd_model",)
54 | FUNCTION = "load_model"
55 | CATEGORY = "DSD"
56 |
57 | def load_model(self, model_path, lora_path, device, dtype, low_cpu_mem_usage, model_cpu_offload, sequential_cpu_offload):
58 | if not IMPORTS_AVAILABLE:
59 | raise ImportError("Could not import DSD modules. Make sure DSD project files (pipeline.py, transformer.py) are properly installed in the parent directory.")
60 |
61 | # Check if model_path is empty, use default path
62 | if not model_path:
63 | model_path = os.path.join(dsd_model_path, "transformer", "diffusion_pytorch_model.safetensors")
64 | print(f"Using default model path: {model_path}")
65 |
66 | # Check if lora_path is empty, use default path
67 | if not lora_path:
68 | lora_path = os.path.join(dsd_model_path, "pytorch_lora_weights.safetensors")
69 | print(f"Using default lora path: {lora_path}")
70 |
71 | # Check if files exist
72 | if not os.path.exists(model_path):
73 | raise FileNotFoundError(f"Model file not found at {model_path}. Please use DSDModelDownloader to download the model first.")
74 |
75 | if not os.path.exists(lora_path):
76 | raise FileNotFoundError(f"LoRA file not found at {lora_path}. Please use DSDModelDownloader to download the model first.")
77 |
78 | print("Loading model...")
79 | # Convert dtype string to torch dtype
80 | torch_dtype = {
81 | "bfloat16": torch.bfloat16,
82 | "float16": torch.float16,
83 | "float32": torch.float32
84 | }.get(dtype, torch.bfloat16)
85 |
86 | print("Loading transformer...")
87 |
88 | model_folder = os.path.dirname(model_path)
89 | # Load model with user-specified parameters
90 | transformer = FluxTransformer2DConditionalModel.from_pretrained(
91 | model_folder,
92 | torch_dtype=torch_dtype,
93 | low_cpu_mem_usage=low_cpu_mem_usage,
94 | ignore_mismatched_sizes=True,
95 | use_safetensors=True,
96 | )
97 |
98 |
99 |
100 | print("Loading pipeline...")
101 |
102 | # Use the optimized from_pretrained method (which was monkey-patched in pipeline.py)
103 | # Use the optimized from_pretrained method
104 | pipe = FluxConditionalPipeline.from_pretrained(
105 | "black-forest-labs/FLUX.1-schnell",
106 | transformer=transformer,
107 | torch_dtype=torch_dtype
108 | )
109 |
110 | # Access and modify scheduler configs
111 | pipe.scheduler.config.shift = 3
112 | pipe.scheduler.config.use_dynamic_shifting = True
113 |
114 | print("Loading LoRA weights...")
115 |
116 | # Load LoRA weights if provided
117 | if lora_path and os.path.exists(lora_path):
118 | pipe.load_lora_weights(lora_path)
119 |
120 | print("Moving to device...")
121 |
122 | # Apply sequential CPU offloading if requested and device is CPU
123 | if model_cpu_offload:
124 | pipe.enable_model_cpu_offload()
125 | if sequential_cpu_offload:
126 | pipe.enable_sequential_cpu_offload()
127 | if not model_cpu_offload and not sequential_cpu_offload:
128 | pipe.to(device)
129 |
130 |
131 |
132 |
133 | print("Model loaded successfully")
134 |
135 | return (pipe,)
136 |
137 |
138 | class DSDGeminiPromptEnhancer:
139 | """Enhances prompts using Google's Gemini API"""
140 |
141 | @classmethod
142 | def INPUT_TYPES(cls):
143 | return {
144 | "required": {
145 | "image": ("IMAGE",),
146 | "prompt": ("STRING", {"multiline": True}),
147 | "api_key": ("STRING", {"default": "", "multiline": False,"tooltip":"Enter your Gemini API key here or use the environment variable GEMINI_API_KEY."})
148 | }
149 | }
150 |
151 | RETURN_TYPES = ("STRING",)
152 | RETURN_NAMES = ("enhanced_prompt",)
153 | FUNCTION = "enhance_prompt"
154 | CATEGORY = "DSD"
155 | OUTPUT_NODE = True # This ensures that UI data is sent to the node
156 |
157 | def __init__(self):
158 | self.enhanced_prompt = None
159 |
160 | def get_state(self):
161 | return {
162 | "enhanced_prompt": self.enhanced_prompt
163 | }
164 |
165 | def enhance_prompt(self, image, prompt, api_key):
166 | if not IMPORTS_AVAILABLE:
167 | print("Warning: DSD modules not available. Using original prompt.")
168 | self.enhanced_prompt = None
169 | return (prompt,)
170 |
171 | if not GEMINI_AVAILABLE:
172 | print("Warning: Google Gemini API not available. Returning original prompt.")
173 | self.enhanced_prompt = None
174 | return (prompt,)
175 |
176 | if not api_key:
177 | #try to get api key from environment variable
178 | api_key = os.getenv("GEMINI_API_KEY")
179 | if not api_key:
180 | print("Warning: No API key provided for Gemini. Returning original prompt.")
181 | self.enhanced_prompt = None
182 | return (prompt,)
183 |
184 | # Convert from ComfyUI image to PIL
185 | pil_image = comfy_to_pil(image)
186 |
187 | # Use the imported enhance_prompt function
188 | try:
189 |
190 | # Call the imported enhance_prompt function
191 | enhanced_prompt = enhance_prompt(pil_image, prompt, api_key)
192 |
193 | print("Original prompt:", prompt)
194 | print("Enhanced prompt:", enhanced_prompt)
195 |
196 | # Store the enhanced prompt for UI display
197 | self.enhanced_prompt = enhanced_prompt
198 |
199 | # Return the enhanced prompt and explicitly include it in the UI data
200 | # Make sure enhanced_prompt is a proper string, not an array/list of characters
201 | return {"ui": {"enhanced_prompt": str(enhanced_prompt)}, "result": (enhanced_prompt,)}
202 | except Exception as e:
203 | print(f"Error enhancing prompt: {e}")
204 | self.enhanced_prompt = None
205 | return (prompt,)
206 |
207 |
208 | class DSDImageGenerator:
209 | """Generates images using the DSD model"""
210 |
211 | @classmethod
212 | def INPUT_TYPES(cls):
213 | return {
214 | "required": {
215 | "dsd_model": ("DSD_MODEL",),
216 | "image": ("IMAGE",),
217 | "prompt": ("STRING", {"multiline": True}),
218 | "negative_prompt": ("STRING", {"multiline": True, "default": ""}),
219 | "seed": ("INT", {"default": 0, "min": 0, "max": 2147483647}),
220 | "guidance_scale": ("FLOAT", {"default": 3.5, "min": 0.0, "max": 20.0, "step": 0.1}),
221 | "image_guidance_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 20.0, "step": 0.1}),
222 | "text_guidance_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 20.0, "step": 0.1}),
223 | "num_inference_steps": ("INT", {"default": 28, "min": 1, "max": 100}),
224 | "width": ("INT", {"default": 1024, "min": 512, "max": 2048, "step": 64}),
225 | "height": ("INT", {"default": 512, "min": 512, "max": 2048, "step": 64}),
226 | "use_gemini_prompt": ("BOOLEAN", {"default": False})
227 | },
228 | "optional": {
229 | "resize_params": ("RESIZE_PARAMS",),
230 | }
231 | }
232 |
233 | RETURN_TYPES = ("IMAGE", "IMAGE", "INT")
234 | RETURN_NAMES = ("image", "reference_image", "seed")
235 | FUNCTION = "generate"
236 | CATEGORY = "DSD"
237 |
238 | def __init__(self):
239 | self.enhanced_prompt = None
240 | # For state tracking
241 | self.progress_value = 0.0
242 |
243 | def get_state(self):
244 | return {
245 | "progress": self.progress_value,
246 | "status_text": self.status_text
247 | }
248 |
249 | @property
250 | def status_text(self):
251 | if self.enhanced_prompt:
252 | return f"Enhanced prompt: {self.enhanced_prompt}"
253 | return ""
254 |
255 | def generate(self, dsd_model, image, prompt, negative_prompt, seed,
256 | guidance_scale, image_guidance_scale, text_guidance_scale, num_inference_steps,
257 | width, height, use_gemini_prompt, resize_params=None):
258 | # Initialize progress bar
259 | pbar = ProgressBar(num_inference_steps)
260 | # Reset progress value
261 | self.progress_value = 0.0
262 |
263 | # Convert from ComfyUI image format to PIL
264 | pil_image = comfy_to_pil(image)
265 |
266 | # Process the image based on resize_params
267 | if resize_params is not None:
268 | method = resize_params.get("method", "center_crop")
269 | interpolation = resize_params.get("interpolation", "LANCZOS")
270 | pad_color = resize_params.get("pad_color", (0, 0, 0))
271 |
272 | # Apply the selected resize method
273 | if method == "resize_and_center_crop":
274 | pil_image = resize_and_center_crop(pil_image, height, width//2)
275 | elif method == "center_crop":
276 | pil_image = center_crop(pil_image, height, width//2, interpolation)
277 | elif method == "pad":
278 | pil_image = pad_resize(pil_image, height, width//2, pad_color, interpolation)
279 | elif method == "fit":
280 | pil_image = fit_resize(pil_image, height, width//2, interpolation)
281 | else:
282 | # Use the default center_crop_and_resize if no resize_params provided
283 | pil_image = resize_and_center_crop(pil_image, height, width//2)
284 |
285 | # Clean prompt
286 | prompt = prompt.strip().replace("\n", "").replace("\r", "")
287 | negative_prompt = negative_prompt.strip().replace("\n", "").replace("\r", "")
288 |
289 | # If use_gemini_prompt enabled, we've already enhanced the prompt, so just store it to show in the UI
290 | if use_gemini_prompt:
291 | try:
292 | self.enhanced_prompt = prompt
293 | except Exception as e:
294 | print(f"Error enhancing prompt: {e}")
295 | self.enhanced_prompt = None
296 | else:
297 | self.enhanced_prompt = None
298 |
299 | # Set up progress callback
300 | def progress_callback(pipe, step, t, callback_kwargs):
301 | # Update the progress bar
302 | pbar.update_absolute(step + 1)
303 | # Update progress value for state tracking
304 | self.progress_value = (step + 1) / num_inference_steps
305 | return callback_kwargs
306 |
307 | # Set up generator with seed
308 | if seed == 0:
309 | # Use a random seed if 0 is provided
310 | seed = torch.randint(0, 2147483647, (1,)).item()
311 | print(f"Using random seed: {seed}")
312 |
313 | # Create generator with the seed
314 | generator = torch.Generator(device="cuda" if torch.cuda.is_available() else "cpu")
315 | generator.manual_seed(seed)
316 |
317 |
318 |
319 | # Run generation
320 | result = dsd_model(
321 | prompt=prompt,
322 | negative_prompt=negative_prompt,
323 | num_inference_steps=num_inference_steps,
324 | height=height,
325 | width=width,
326 | guidance_scale=guidance_scale,
327 | image=pil_image,
328 | guidance_scale_real_i=image_guidance_scale,
329 | guidance_scale_real_t=text_guidance_scale,
330 | callback_on_step_end=progress_callback,
331 | generator=generator
332 | ).images
333 |
334 | # Debug information
335 | print(f"Result type: {type(result)}")
336 | print(f"Result length: {len(result)}")
337 |
338 | # Get the output image (right side)
339 | output_image = None
340 | if result[0] is not None:
341 | print(f"Output image type: {type(result[0])}")
342 | if isinstance(result[0], list):
343 | print(f"Output image list length: {len(result[0])}")
344 | if len(result[0]) > 0:
345 | print(f"First output image type: {type(result[0][0])}, size: {result[0][0].size if hasattr(result[0][0], 'size') else 'unknown'}")
346 | else:
347 | print(f"Output image size: {result[0].size if hasattr(result[0], 'size') else 'unknown'}")
348 |
349 | # Convert to ComfyUI format
350 | output_image = pil_to_comfy(result[0])
351 | if output_image is not None:
352 | print(f"Converted output image shape: {output_image.shape}")
353 |
354 | # Get the reference image (left side)
355 | reference_image = None
356 | if len(result) > 1 and result[1] is not None:
357 | print(f"Reference image type: {type(result[1])}")
358 | if isinstance(result[1], list):
359 | print(f"Reference image list length: {len(result[1])}")
360 | if len(result[1]) > 0:
361 | print(f"First reference image type: {type(result[1][0])}, size: {result[1][0].size if hasattr(result[1][0], 'size') else 'unknown'}")
362 | else:
363 | print(f"Reference image size: {result[1].size if hasattr(result[1], 'size') else 'unknown'}")
364 |
365 | # Convert to ComfyUI format
366 | reference_image = pil_to_comfy(result[1])
367 | if reference_image is not None:
368 | print(f"Converted reference image shape: {reference_image.shape}")
369 |
370 | # Ensure progress is complete
371 | pbar.update_absolute(num_inference_steps)
372 | self.progress_value = 1.0
373 |
374 |
375 |
376 | return (output_image, reference_image, seed)
377 |
378 |
379 | # Add a separate model selector that helps find models in the model directory
380 | class DSDModelSelector:
381 | """Selects a DSD model from the models directory"""
382 |
383 | @classmethod
384 | def INPUT_TYPES(cls):
385 | return {
386 | "required": {
387 | "check_model_exists": ("BOOLEAN", {"default": True}),
388 | }
389 | }
390 |
391 | RETURN_TYPES = ("STRING", "STRING")
392 | RETURN_NAMES = ("model_path", "lora_path")
393 | FUNCTION = "select_model"
394 | CATEGORY = "DSD"
395 |
396 | def select_model(self, check_model_exists):
397 | # Get the transformer path
398 | model_path = os.path.join(dsd_model_path, "transformer", "diffusion_pytorch_model.safetensors")
399 | # Get the lora path
400 | lora_path = os.path.join(dsd_model_path, "pytorch_lora_weights.safetensors")
401 |
402 | # Check if files exist
403 | if check_model_exists:
404 | if not os.path.exists(model_path):
405 | print(f"Warning: Model file not found at {model_path}")
406 | print("You can use DSDModelDownloader to download and load the model.")
407 |
408 | if not os.path.exists(lora_path):
409 | print(f"Warning: LoRA file not found at {lora_path}")
410 | print("You can use DSDModelDownloader to download and load the model.")
411 |
412 | return (model_path, lora_path)
413 |
414 |
415 | class DSDModelDownloader:
416 | """Downloads and loads the DSD model from Hugging Face"""
417 |
418 | @classmethod
419 | def INPUT_TYPES(cls):
420 | return {
421 | "required": {
422 | "repo_id": ("STRING", {"default": "primecai/dsd_model"}),
423 | "force_download": ("BOOLEAN", {"default": False}),
424 | "device": (["cuda", "cpu"], {"default": "cuda"}),
425 | "dtype": (["bfloat16", "float16", "float32"], {"default": "bfloat16", "tooltip": "bfloat16 provides best speed/memory tradeoff"}),
426 | "low_cpu_mem_usage": ("BOOLEAN", {"default": True, "tooltip": "Reduces CPU memory usage during model loading. Recommended for faster loading."}),
427 | "model_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "Offloads state dict to reduce memory usage during loading. May slow down loading speed."}),
428 | "sequential_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "Enables sequential CPU offloading. Only use if low on VRAM. Significantly impacts loading speed."})
429 | }
430 | }
431 |
432 | RETURN_TYPES = ("DSD_MODEL", "STRING", "STRING")
433 | RETURN_NAMES = ("dsd_model", "model_path", "lora_path")
434 | FUNCTION = "download_and_load_model"
435 | CATEGORY = "DSD"
436 |
437 | def __init__(self):
438 | self.progress_value = 0.0
439 | self.status_text = ""
440 |
441 | def get_state(self):
442 | return {
443 | "progress": self.progress_value,
444 | "status_text": self.status_text
445 | }
446 |
447 | def download_and_load_model(self, repo_id, force_download, device, dtype, low_cpu_mem_usage, model_cpu_offload, sequential_cpu_offload):
448 | if not IMPORTS_AVAILABLE:
449 | raise ImportError("Could not import DSD modules. Make sure DSD project files (pipeline.py, transformer.py) are properly installed in the parent directory.")
450 |
451 | # Create the dsd_model directory in ComfyUI models folder if it doesn't exist
452 | os.makedirs(dsd_model_path, exist_ok=True)
453 | transformer_path = os.path.join(dsd_model_path, "transformer")
454 | os.makedirs(transformer_path, exist_ok=True)
455 |
456 | # Check if model already exists
457 | model_file = os.path.join(transformer_path, "diffusion_pytorch_model.safetensors")
458 | config_file = os.path.join(transformer_path, "config.json")
459 | lora_file = os.path.join(dsd_model_path, "pytorch_lora_weights.safetensors")
460 |
461 | files_exist = os.path.exists(model_file) and os.path.exists(config_file) and os.path.exists(lora_file)
462 |
463 | if not files_exist or force_download:
464 | self.status_text = f"Downloading DSD model from {repo_id}..."
465 | print(self.status_text)
466 | self.progress_value = 0.1
467 |
468 | try:
469 | # Download the model files
470 | self.status_text = "Downloading transformer model and LoRA weights..."
471 | print(self.status_text)
472 |
473 | # Use snapshot_download to download all files
474 | snapshot_download(
475 | repo_id=repo_id,
476 | local_dir=dsd_model_path,
477 | local_dir_use_symlinks=False,
478 | resume_download=True
479 | )
480 |
481 | self.progress_value = 0.5
482 | self.status_text = "Model downloaded successfully"
483 | print(self.status_text)
484 |
485 | # Verify files were downloaded correctly
486 | if not os.path.exists(model_file):
487 | raise FileNotFoundError(f"Model file not found at {model_file} after download. The repository structure may be different than expected.")
488 |
489 | if not os.path.exists(config_file):
490 | raise FileNotFoundError(f"Config file not found at {config_file} after download. The repository structure may be different than expected.")
491 |
492 | if not os.path.exists(lora_file):
493 | raise FileNotFoundError(f"LoRA file not found at {lora_file} after download. The repository structure may be different than expected.")
494 |
495 | except Exception as e:
496 | self.status_text = f"Error downloading model: {str(e)}"
497 | print(self.status_text)
498 | raise
499 | else:
500 | self.status_text = "Model files already exist. Skipping download."
501 | print(self.status_text)
502 | self.progress_value = 0.5
503 |
504 | # Convert dtype string to torch dtype
505 | torch_dtype = {
506 | "bfloat16": torch.bfloat16,
507 | "float16": torch.float16,
508 | "float32": torch.float32
509 | }.get(dtype, torch.bfloat16)
510 |
511 | self.status_text = "Loading transformer..."
512 | print(self.status_text)
513 | self.progress_value = 0.6
514 |
515 | try:
516 | # Load model with user-specified parameters
517 | transformer = FluxTransformer2DConditionalModel.from_pretrained(
518 | transformer_path,
519 | torch_dtype=torch_dtype,
520 | low_cpu_mem_usage=low_cpu_mem_usage,
521 | ignore_mismatched_sizes=True,
522 | use_safetensors=True
523 | )
524 |
525 | self.status_text = "Loading pipeline..."
526 | print(self.status_text)
527 | self.progress_value = 0.7
528 |
529 |
530 |
531 |
532 | # Using black-forest-labs/FLUX.1-schnell,so we don't need to login to Hugging Face
533 | pipe = FluxConditionalPipeline.from_pretrained(
534 | "black-forest-labs/FLUX.1-schnell",
535 | transformer=transformer,
536 | torch_dtype=torch_dtype
537 | )
538 |
539 | # Access and modify scheduler configs
540 | pipe.scheduler.config.shift = 3
541 | pipe.scheduler.config.use_dynamic_shifting = True
542 |
543 |
544 |
545 | self.status_text = "Loading LoRA weights..."
546 | print(self.status_text)
547 | self.progress_value = 0.8
548 |
549 | # Load LoRA weights
550 | pipe.load_lora_weights(lora_file)
551 |
552 | # Apply sequential CPU offloading if requested and device is CPU
553 | if model_cpu_offload:
554 | pipe.enable_model_cpu_offload()
555 | if sequential_cpu_offload:
556 | pipe.enable_sequential_cpu_offload()
557 | if not model_cpu_offload and not sequential_cpu_offload:
558 | pipe.to(device)
559 |
560 | self.progress_value = 0.9
561 |
562 | self.status_text = "Model loaded successfully"
563 | print(self.status_text)
564 | self.progress_value = 1.0
565 |
566 | return (pipe, model_file, lora_file)
567 |
568 | except Exception as e:
569 | self.status_text = f"Error loading model: {str(e)}"
570 | print(self.status_text)
571 | raise
572 |
573 |
574 | class DSDResizeSelector:
575 | """Selects image resize options for DSD Image Generator"""
576 |
577 | @classmethod
578 | def INPUT_TYPES(cls):
579 | return {
580 | "required": {
581 | "resize_method": (["resize_and_center_crop", "center_crop", "pad", "fit"], {"default": "resize_and_center_crop"}),
582 | "interpolation": (["LANCZOS", "BICUBIC", "BILINEAR", "NEAREST"], {"default": "LANCZOS"}),
583 | "pad_r": ("INT", {"default": 0, "min": 0, "max": 255}),
584 | "pad_g": ("INT", {"default": 0, "min": 0, "max": 255}),
585 | "pad_b": ("INT", {"default": 0, "min": 0, "max": 255}),
586 | }
587 | }
588 |
589 | RETURN_TYPES = ("RESIZE_PARAMS",)
590 | RETURN_NAMES = ("resize_params",)
591 | FUNCTION = "select_resize_options"
592 | CATEGORY = "DSD"
593 |
594 | def select_resize_options(self, resize_method, interpolation, pad_r, pad_g, pad_b):
595 | # Create a JSON object with the resize parameters
596 | resize_params = {
597 | "method": resize_method,
598 | "interpolation": interpolation,
599 | "pad_color": (pad_r, pad_g, pad_b)
600 | }
601 |
602 | return (resize_params,)
603 |
604 |
605 | # Register nodes
606 | NODE_CLASS_MAPPINGS = {
607 | "DSDModelLoader": DSDModelLoader,
608 | "DSDGeminiPromptEnhancer": DSDGeminiPromptEnhancer,
609 | "DSDImageGenerator": DSDImageGenerator,
610 | "DSDModelSelector": DSDModelSelector,
611 | "DSDModelDownloader": DSDModelDownloader,
612 | "DSDResizeSelector": DSDResizeSelector
613 | }
614 |
615 | # Node display names
616 | NODE_DISPLAY_NAME_MAPPINGS = {
617 | "DSDModelLoader": "DSD Model Loader",
618 | "DSDGeminiPromptEnhancer": "DSD Gemini Prompt Enhancer",
619 | "DSDImageGenerator": "DSD Image Generator",
620 | "DSDModelSelector": "DSD Model Selector",
621 | "DSDModelDownloader": "DSD Model Downloader",
622 | "DSDResizeSelector": "DSD Resize Selector"
623 | }
--------------------------------------------------------------------------------
/examples/example_workflow.json:
--------------------------------------------------------------------------------
1 | {
2 | "last_node_id": 39,
3 | "last_link_id": 117,
4 | "nodes": [
5 | {
6 | "id": 18,
7 | "type": "PreviewImage",
8 | "pos": [
9 | 1214.10986328125,
10 | -74.78528594970703
11 | ],
12 | "size": [
13 | 406.026123046875,
14 | 409.0712585449219
15 | ],
16 | "flags": {},
17 | "order": 7,
18 | "mode": 0,
19 | "inputs": [
20 | {
21 | "name": "images",
22 | "type": "IMAGE",
23 | "link": 115
24 | }
25 | ],
26 | "outputs": [],
27 | "properties": {
28 | "cnr_id": "comfy-core",
29 | "ver": "0.3.26",
30 | "Node name for S&R": "PreviewImage"
31 | },
32 | "widgets_values": []
33 | },
34 | {
35 | "id": 3,
36 | "type": "LoadImage",
37 | "pos": [
38 | -119.63543701171875,
39 | -71.78013610839844
40 | ],
41 | "size": [
42 | 366.44219970703125,
43 | 419.337890625
44 | ],
45 | "flags": {},
46 | "order": 0,
47 | "mode": 0,
48 | "inputs": [],
49 | "outputs": [
50 | {
51 | "name": "IMAGE",
52 | "type": "IMAGE",
53 | "shape": 3,
54 | "links": [
55 | 4,
56 | 109
57 | ],
58 | "slot_index": 0
59 | },
60 | {
61 | "name": "MASK",
62 | "type": "MASK",
63 | "shape": 3,
64 | "links": []
65 | }
66 | ],
67 | "properties": {
68 | "cnr_id": "comfy-core",
69 | "ver": "0.3.26",
70 | "Node name for S&R": "LoadImage"
71 | },
72 | "widgets_values": [
73 | "DALL·E 2024-08-18 18.34.08 - A 2D anime-style character concept art in the style of Porco Rosso. The character is a young, male airplane mechanic in his early 20s, with messy brow.webp",
74 | "image"
75 | ]
76 | },
77 | {
78 | "id": 19,
79 | "type": "PreviewImage",
80 | "pos": [
81 | 1271.379638671875,
82 | 431.7550964355469
83 | ],
84 | "size": [
85 | 210,
86 | 246
87 | ],
88 | "flags": {},
89 | "order": 8,
90 | "mode": 0,
91 | "inputs": [
92 | {
93 | "name": "images",
94 | "type": "IMAGE",
95 | "link": 113
96 | }
97 | ],
98 | "outputs": [],
99 | "properties": {
100 | "cnr_id": "comfy-core",
101 | "ver": "0.3.26",
102 | "Node name for S&R": "PreviewImage"
103 | },
104 | "widgets_values": []
105 | },
106 | {
107 | "id": 5,
108 | "type": "DSDGeminiPromptEnhancer",
109 | "pos": [
110 | 299.41937255859375,
111 | 386.8332824707031
112 | ],
113 | "size": [
114 | 334.03948974609375,
115 | 137.88925170898438
116 | ],
117 | "flags": {},
118 | "order": 5,
119 | "mode": 0,
120 | "inputs": [
121 | {
122 | "name": "image",
123 | "type": "IMAGE",
124 | "link": 4
125 | },
126 | {
127 | "name": "prompt",
128 | "type": "STRING",
129 | "widget": {
130 | "name": "prompt"
131 | },
132 | "link": 5
133 | }
134 | ],
135 | "outputs": [
136 | {
137 | "name": "enhanced_prompt",
138 | "type": "STRING",
139 | "shape": 3,
140 | "links": [
141 | 110
142 | ],
143 | "slot_index": 0
144 | }
145 | ],
146 | "properties": {
147 | "cnr_id": "dsd",
148 | "ver": "4642def54ab46095a128cc2f8d37abde99a0f099",
149 | "Node name for S&R": "DSDGeminiPromptEnhancer",
150 | "aux_id": "irreveloper/ComfyUI-DSD-Node"
151 | },
152 | "widgets_values": [
153 | "Side view of anime character.",
154 | ""
155 | ]
156 | },
157 | {
158 | "id": 39,
159 | "type": "DSDResizeSelector",
160 | "pos": [
161 | 314.8365478515625,
162 | 171.76687622070312
163 | ],
164 | "size": [
165 | 315,
166 | 154
167 | ],
168 | "flags": {},
169 | "order": 1,
170 | "mode": 0,
171 | "inputs": [],
172 | "outputs": [
173 | {
174 | "name": "resize_params",
175 | "type": "RESIZE_PARAMS",
176 | "links": [
177 | 117
178 | ],
179 | "slot_index": 0
180 | }
181 | ],
182 | "properties": {
183 | "cnr_id": "dsd",
184 | "ver": "48d020ac58d05c8a14223dac597a645ee908684a",
185 | "Node name for S&R": "DSDResizeSelector"
186 | },
187 | "widgets_values": [
188 | "resize_and_center_crop",
189 | "LANCZOS",
190 | 0,
191 | 0,
192 | 0
193 | ]
194 | },
195 | {
196 | "id": 38,
197 | "type": "DSDModelDownloader",
198 | "pos": [
199 | 328.3520812988281,
200 | -165.09207153320312
201 | ],
202 | "size": [
203 | 315,
204 | 266
205 | ],
206 | "flags": {},
207 | "order": 2,
208 | "mode": 0,
209 | "inputs": [],
210 | "outputs": [
211 | {
212 | "name": "dsd_model",
213 | "type": "DSD_MODEL",
214 | "links": [
215 | 116
216 | ]
217 | },
218 | {
219 | "name": "model_path",
220 | "type": "STRING",
221 | "links": null
222 | },
223 | {
224 | "name": "lora_path",
225 | "type": "STRING",
226 | "links": null
227 | }
228 | ],
229 | "properties": {
230 | "cnr_id": "dsd",
231 | "ver": "4642def54ab46095a128cc2f8d37abde99a0f099",
232 | "Node name for S&R": "DSDModelDownloader",
233 | "aux_id": "irreveloper/ComfyUI-DSD-Node"
234 | },
235 | "widgets_values": [
236 | "primecai/dsd_model",
237 | false,
238 | "cuda",
239 | "bfloat16",
240 | true,
241 | false,
242 | false,
243 | ""
244 | ]
245 | },
246 | {
247 | "id": 4,
248 | "type": "PrimitiveNode",
249 | "pos": [
250 | -117.78815460205078,
251 | 407.2402648925781
252 | ],
253 | "size": [
254 | 315,
255 | 130
256 | ],
257 | "flags": {},
258 | "order": 3,
259 | "mode": 0,
260 | "inputs": [],
261 | "outputs": [
262 | {
263 | "name": "STRING",
264 | "type": "STRING",
265 | "shape": 3,
266 | "links": [
267 | 5
268 | ],
269 | "slot_index": 0
270 | }
271 | ],
272 | "properties": {
273 | "Run widget replace on values": false
274 | },
275 | "widgets_values": [
276 | "Side view of anime character."
277 | ]
278 | },
279 | {
280 | "id": 11,
281 | "type": "PrimitiveNode",
282 | "pos": [
283 | -106.84088897705078,
284 | 607.5902099609375
285 | ],
286 | "size": [
287 | 301.3577880859375,
288 | 88
289 | ],
290 | "flags": {},
291 | "order": 4,
292 | "mode": 0,
293 | "inputs": [],
294 | "outputs": [
295 | {
296 | "name": "STRING",
297 | "type": "STRING",
298 | "links": [
299 | 111
300 | ]
301 | }
302 | ],
303 | "title": "negative_prompt",
304 | "properties": {
305 | "Run widget replace on values": false
306 | },
307 | "widgets_values": [
308 | "text, watermark, blurry"
309 | ]
310 | },
311 | {
312 | "id": 37,
313 | "type": "DSDImageGenerator",
314 | "pos": [
315 | 721.5396728515625,
316 | 86.6509017944336
317 | ],
318 | "size": [
319 | 400,
320 | 398
321 | ],
322 | "flags": {},
323 | "order": 6,
324 | "mode": 0,
325 | "inputs": [
326 | {
327 | "name": "dsd_model",
328 | "type": "DSD_MODEL",
329 | "link": 116
330 | },
331 | {
332 | "name": "image",
333 | "type": "IMAGE",
334 | "link": 109
335 | },
336 | {
337 | "name": "prompt",
338 | "type": "STRING",
339 | "widget": {
340 | "name": "prompt"
341 | },
342 | "link": 110
343 | },
344 | {
345 | "name": "negative_prompt",
346 | "type": "STRING",
347 | "widget": {
348 | "name": "negative_prompt"
349 | },
350 | "link": 111
351 | },
352 | {
353 | "name": "resize_params",
354 | "type": "RESIZE_PARAMS",
355 | "shape": 7,
356 | "link": 117
357 | }
358 | ],
359 | "outputs": [
360 | {
361 | "name": "image",
362 | "type": "IMAGE",
363 | "links": [
364 | 115
365 | ]
366 | },
367 | {
368 | "name": "reference_image",
369 | "type": "IMAGE",
370 | "links": [
371 | 113
372 | ]
373 | },
374 | {
375 | "name": "seed",
376 | "type": "INT",
377 | "links": []
378 | }
379 | ],
380 | "properties": {
381 | "cnr_id": "dsd",
382 | "ver": "4642def54ab46095a128cc2f8d37abde99a0f099",
383 | "Node name for S&R": "DSDImageGenerator",
384 | "aux_id": "irreveloper/ComfyUI-DSD-Node"
385 | },
386 | "widgets_values": [
387 | "",
388 | "text, watermark, blurry",
389 | 3,
390 | "fixed",
391 | 3.5,
392 | 1,
393 | 1,
394 | 28,
395 | 1024,
396 | 512,
397 | true
398 | ]
399 | }
400 | ],
401 | "links": [
402 | [
403 | 4,
404 | 3,
405 | 0,
406 | 5,
407 | 0,
408 | "IMAGE"
409 | ],
410 | [
411 | 5,
412 | 4,
413 | 0,
414 | 5,
415 | 1,
416 | "STRING"
417 | ],
418 | [
419 | 109,
420 | 3,
421 | 0,
422 | 37,
423 | 1,
424 | "IMAGE"
425 | ],
426 | [
427 | 110,
428 | 5,
429 | 0,
430 | 37,
431 | 2,
432 | "STRING"
433 | ],
434 | [
435 | 111,
436 | 11,
437 | 0,
438 | 37,
439 | 3,
440 | "STRING"
441 | ],
442 | [
443 | 113,
444 | 37,
445 | 1,
446 | 19,
447 | 0,
448 | "IMAGE"
449 | ],
450 | [
451 | 115,
452 | 37,
453 | 0,
454 | 18,
455 | 0,
456 | "IMAGE"
457 | ],
458 | [
459 | 116,
460 | 38,
461 | 0,
462 | 37,
463 | 0,
464 | "DSD_MODEL"
465 | ],
466 | [
467 | 117,
468 | 39,
469 | 0,
470 | 37,
471 | 4,
472 | "RESIZE_PARAMS"
473 | ]
474 | ],
475 | "groups": [],
476 | "config": {},
477 | "extra": {
478 | "ds": {
479 | "scale": 0.8768324088719838,
480 | "offset": [
481 | 773.2185529985686,
482 | 450.555653889974
483 | ]
484 | }
485 | },
486 | "version": 0.4
487 | }
--------------------------------------------------------------------------------
/examples/workflow.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/irreveloper/ComfyUI-DSD/6af416966bdc6373c402079289661f37e8070061/examples/workflow.png
--------------------------------------------------------------------------------
/pyproject.toml:
--------------------------------------------------------------------------------
1 | [project]
2 | name = "dsd"
3 | description = "An Unofficial ComfyUI custom node package that integrates [a/Diffusion Self-Distillation (DSD)](https://github.com/primecai/diffusion-self-distillation) for zero-shot customized image generation.\nDSD is a model for subject-preserving image generation that allows you to create images of a specific subject in novel contexts without per-instance tuning."
4 | version = "1.0.2"
5 | license = {file = "LICENSE"}
6 | dependencies = ["torch>=2.0.0", "diffusers>=0.24.0", "transformers>=4.36.0", "sentencepiece>=0.1.99", "accelerate>=0.27.0", "google-genai>=0.1.0", "Pillow>=9.5.0", "protobuf>=4.25.0", "peft>=0.7.0", "hf_transfer"]
7 |
8 | [project.urls]
9 | Repository = "https://github.com/irreveloper/ComfyUI-DSD"
10 | # Used by Comfy Registry https://comfyregistry.org
11 |
12 | [tool.comfy]
13 | PublisherId = "irreveloper"
14 | DisplayName = "ComfyUI-DSD"
15 | Icon = ""
16 |
--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | torch>=2.0.0
2 | diffusers>=0.24.0
3 | transformers>=4.36.0
4 | sentencepiece>=0.1.99
5 | accelerate>=0.27.0
6 | google-genai>=0.1.0
7 | Pillow>=9.5.0
8 | protobuf>=4.25.0
9 | peft>=0.7.0
10 | hf_transfer
--------------------------------------------------------------------------------
/utils.py:
--------------------------------------------------------------------------------
1 | import os
2 | import torch
3 | import numpy as np
4 | from PIL import Image
5 | from typing import Union, List, Optional
6 |
7 | def get_model_path(model_name: str) -> str:
8 | """
9 | Get the path to a model file in the models directory.
10 |
11 | Args:
12 | model_name: Name of the model file or directory
13 |
14 | Returns:
15 | Full path to the model
16 | """
17 | # Determine the path to the models directory
18 | current_dir = os.path.dirname(os.path.abspath(__file__))
19 | models_dir = os.path.join(current_dir, "models")
20 |
21 | # Check if the model exists
22 | model_path = os.path.join(models_dir, model_name)
23 | if not os.path.exists(model_path):
24 | raise FileNotFoundError(f"Model {model_name} not found in {models_dir}")
25 |
26 | return model_path
27 |
28 | def get_lora_path(lora_name: str) -> str:
29 | """
30 | Get the path to a LoRA file in the loras directory.
31 |
32 | Args:
33 | lora_name: Name of the LoRA file
34 |
35 | Returns:
36 | Full path to the LoRA file
37 | """
38 | # Determine the path to the loras directory
39 | current_dir = os.path.dirname(os.path.abspath(__file__))
40 | loras_dir = os.path.join(current_dir, "loras")
41 |
42 | # Check if the LoRA file exists
43 | lora_path = os.path.join(loras_dir, lora_name)
44 | if not os.path.exists(lora_path):
45 | raise FileNotFoundError(f"LoRA file {lora_name} not found in {loras_dir}")
46 |
47 | return lora_path
48 |
49 | def comfy_to_pil(image: torch.Tensor) -> Image.Image:
50 | """
51 | Convert a ComfyUI image tensor to a PIL Image.
52 |
53 | Args:
54 | image: ComfyUI image tensor (1, H, W, 3) in range [0, 1]
55 |
56 | Returns:
57 | PIL Image
58 | """
59 | # Convert to numpy array and scale to [0, 255]
60 | image_np = np.clip(255. * image[0].cpu().numpy(), 0, 255).astype(np.uint8)
61 | # Convert to PIL Image
62 | return Image.fromarray(image_np)
63 |
64 | def pil_to_comfy(image: Union[Image.Image, List[Image.Image], None]) -> Optional[torch.Tensor]:
65 | """
66 | Convert a PIL Image or list of PIL Images to a ComfyUI image tensor.
67 |
68 | Args:
69 | image: PIL Image, list of PIL Images, or None
70 |
71 | Returns:
72 | ComfyUI image tensor (1, H, W, 3) in range [0, 1] or None if input is None
73 | """
74 | if image is None:
75 | return None
76 |
77 | # Handle list of PIL images - take the first one
78 | if isinstance(image, list):
79 | if len(image) == 0:
80 | return None
81 | image = image[0] # Take the first image from the list
82 |
83 | # Convert to numpy array and scale to [0, 1]
84 | image_np = np.array(image).astype(np.float32) / 255.0
85 |
86 | # Ensure the image has the right shape (H, W, 3)
87 | if len(image_np.shape) == 2: # Grayscale image
88 | image_np = np.stack([image_np, image_np, image_np], axis=-1)
89 | elif image_np.shape[-1] == 4: # RGBA image
90 | image_np = image_np[..., :3] # Remove alpha channel
91 |
92 | # Convert to torch tensor and add batch dimension
93 | return torch.from_numpy(image_np)[None,]
94 |
95 | def resize_and_center_crop(image: Union[torch.Tensor, Image.Image], target_height: int = 512,target_width: int = 512) -> Union[torch.Tensor, Image.Image]:
96 | """
97 | Center crop and resize an image.
98 |
99 | Args:
100 | image: Image to process (ComfyUI tensor or PIL Image)
101 | target_height: Target height
102 | target_width: Target width
103 |
104 | Returns:
105 | Processed image in the same format as input
106 | """
107 | # Handle ComfyUI tensor
108 | if isinstance(image, torch.Tensor):
109 | pil_image = comfy_to_pil(image)
110 | result = resize_and_center_crop(pil_image, target_height, target_width)
111 | return pil_to_comfy(result)
112 |
113 | # Handle PIL Image
114 | w, h = image.size
115 | min_size = min(w, h)
116 |
117 | # Calculate target aspect ratio
118 | target_ratio = target_width / target_height
119 | # Calculate current aspect ratio
120 | current_ratio = w / h
121 |
122 | # Resize to match target width or height while preserving aspect ratio
123 | if current_ratio > target_ratio:
124 | # Image is wider than target - resize by height
125 | new_height = target_height
126 | new_width = int(w * (target_height / h))
127 | else:
128 | # Image is taller than target - resize by width
129 | new_width = target_width
130 | new_height = int(h * (target_width / w))
131 |
132 | image = image.resize((new_width, new_height), Image.BILINEAR)
133 |
134 | # Center crop the image to the target size
135 | cropped = image.crop(((new_width - target_width) // 2,
136 | (new_height - target_height) // 2,
137 | (new_width + target_width) // 2,
138 | (new_height + target_height) // 2))
139 |
140 | return cropped
141 |
142 | def center_crop(image: Union[torch.Tensor, Image.Image], target_height: int = 512, target_width: int = 512,
143 | interpolation: str = "LANCZOS") -> Union[torch.Tensor, Image.Image]:
144 | """
145 | Center crop an image to target size. If image is smaller than target size,
146 | resize first before cropping.
147 |
148 | Args:
149 | image: Image to process (ComfyUI tensor or PIL Image)
150 | target_height: Target height
151 | target_width: Target width
152 | interpolation: Interpolation method (NEAREST, BILINEAR, BICUBIC, LANCZOS)
153 |
154 | Returns:
155 | Processed image in the same format as input
156 | """
157 | # Handle ComfyUI tensor
158 | if isinstance(image, torch.Tensor):
159 | pil_image = comfy_to_pil(image)
160 | result = center_crop(pil_image, target_height, target_width, interpolation)
161 | return pil_to_comfy(result)
162 |
163 | # Handle PIL Image
164 | w, h = image.size
165 |
166 | # Get interpolation method
167 | interp_method = {
168 | "NEAREST": Image.NEAREST,
169 | "BILINEAR": Image.BILINEAR,
170 | "BICUBIC": Image.BICUBIC,
171 | "LANCZOS": Image.LANCZOS
172 | }.get(interpolation, Image.LANCZOS)
173 |
174 | # If image is smaller than target in either dimension, resize first
175 | if w < target_width or h < target_height:
176 | # Calculate scale factor needed
177 | scale = max(target_width / w, target_height / h)
178 | new_w = int(w * scale)
179 | new_h = int(h * scale)
180 | image = image.resize((new_w, new_h), interp_method)
181 | w, h = new_w, new_h
182 |
183 | # Calculate crop coordinates
184 | left = (w - target_width) // 2
185 | top = (h - target_height) // 2
186 | right = left + target_width
187 | bottom = top + target_height
188 |
189 | # Perform center crop
190 | cropped = image.crop((left, top, right, bottom))
191 | return cropped
192 |
193 | def pad_resize(image: Union[torch.Tensor, Image.Image], target_height: int = 512, target_width: int = 512,
194 | pad_color: tuple = (0, 0, 0), interpolation: str = "LANCZOS") -> Union[torch.Tensor, Image.Image]:
195 | """
196 | Resize image preserving aspect ratio and pad to target size.
197 |
198 | Args:
199 | image: Image to process (ComfyUI tensor or PIL Image)
200 | target_height: Target height
201 | target_width: Target width
202 | pad_color: RGB color tuple for padding (default: black)
203 | interpolation: Interpolation method (NEAREST, BILINEAR, BICUBIC, LANCZOS)
204 |
205 | Returns:
206 | Processed image in the same format as input
207 | """
208 | # Handle ComfyUI tensor
209 | if isinstance(image, torch.Tensor):
210 | pil_image = comfy_to_pil(image)
211 | result = pad_resize(pil_image, target_height, target_width, pad_color, interpolation)
212 | return pil_to_comfy(result)
213 |
214 | # Handle PIL Image
215 | w, h = image.size
216 |
217 | # Get interpolation method
218 | interp_method = {
219 | "NEAREST": Image.NEAREST,
220 | "BILINEAR": Image.BILINEAR,
221 | "BICUBIC": Image.BICUBIC,
222 | "LANCZOS": Image.LANCZOS
223 | }.get(interpolation, Image.LANCZOS)
224 |
225 | # Calculate target aspect ratio
226 | target_ratio = target_width / target_height
227 | # Calculate current aspect ratio
228 | current_ratio = w / h
229 |
230 | # Create a new image with the target size and fill with the pad color
231 | new_image = Image.new("RGB", (target_width, target_height), pad_color)
232 |
233 | # Resize the original image preserving aspect ratio
234 | if current_ratio > target_ratio:
235 | # Image is wider than target - resize by width
236 | new_w = target_width
237 | new_h = int(h * (target_width / w))
238 | resized = image.resize((new_w, new_h), interp_method)
239 | # Paste in the center (horizontally)
240 | new_image.paste(resized, (0, (target_height - new_h) // 2))
241 | else:
242 | # Image is taller than target - resize by height
243 | new_h = target_height
244 | new_w = int(w * (target_height / h))
245 | resized = image.resize((new_w, new_h), interp_method)
246 | # Paste in the center (vertically)
247 | new_image.paste(resized, ((target_width - new_w) // 2, 0))
248 |
249 | return new_image
250 |
251 | def fit_resize(image: Union[torch.Tensor, Image.Image], target_height: int = 512, target_width: int = 512,
252 | interpolation: str = "LANCZOS") -> Union[torch.Tensor, Image.Image]:
253 | """
254 | Resize image to target size without preserving aspect ratio.
255 |
256 | Args:
257 | image: Image to process (ComfyUI tensor or PIL Image)
258 | target_height: Target height
259 | target_width: Target width
260 | interpolation: Interpolation method (NEAREST, BILINEAR, BICUBIC, LANCZOS)
261 |
262 | Returns:
263 | Processed image in the same format as input
264 | """
265 | # Handle ComfyUI tensor
266 | if isinstance(image, torch.Tensor):
267 | pil_image = comfy_to_pil(image)
268 | result = fit_resize(pil_image, target_height, target_width, interpolation)
269 | return pil_to_comfy(result)
270 |
271 | # Handle PIL Image
272 | # Get interpolation method
273 | interp_method = {
274 | "NEAREST": Image.NEAREST,
275 | "BILINEAR": Image.BILINEAR,
276 | "BICUBIC": Image.BICUBIC,
277 | "LANCZOS": Image.LANCZOS
278 | }.get(interpolation, Image.LANCZOS)
279 |
280 | # Resize directly to target size (no aspect ratio preservation)
281 | resized = image.resize((target_width, target_height), interp_method)
282 | return resized
--------------------------------------------------------------------------------
/web/js/showEnhancedPrompt.js:
--------------------------------------------------------------------------------
1 | import { app } from "../../../scripts/app.js";
2 | import { ComfyWidgets } from "../../../scripts/widgets.js";
3 |
4 | // Displays enhanced prompt on DSDGeminiPromptEnhancer node
5 | app.registerExtension({
6 | name: "comfyui.dsd.showEnhancedPrompt",
7 | async beforeRegisterNodeDef(nodeType, nodeData, app) {
8 | if (nodeData.name === "DSDGeminiPromptEnhancer") {
9 | function normalizePrompt(prompt) {
10 | if (!prompt) return "";
11 | if (Array.isArray(prompt)) {
12 | return prompt.join("");
13 | }
14 | return prompt;
15 | }
16 |
17 | function populateEnhancedPrompt(enhancedPrompt) {
18 | enhancedPrompt = normalizePrompt(enhancedPrompt);
19 | if (!enhancedPrompt) {
20 | return;
21 | }
22 |
23 | try {
24 | if (this.widgets) {
25 | for (let i = 0; i < this.widgets.length; i++) {
26 | if (this.widgets[i].name === "enhanced_prompt_display") {
27 | this.widgets[i].onRemove?.();
28 | this.widgets.splice(i, 1);
29 | i--;
30 | }
31 | }
32 | }
33 |
34 | const textWidgetResult = ComfyWidgets["STRING"](this, "enhanced_prompt_text", ["STRING", { multiline: true }], app);
35 | if (!textWidgetResult || !textWidgetResult.widget || !textWidgetResult.widget.inputEl) {
36 | return;
37 | }
38 | const w = textWidgetResult.widget;
39 | w.inputEl.readOnly = true;
40 | w.inputEl.style.opacity = 0.8;
41 | w.inputEl.style.backgroundColor = "#1e2124";
42 | w.inputEl.style.color = "#9eec51";
43 | w.name = "enhanced_prompt_display";
44 | w.value = enhancedPrompt;
45 |
46 | requestAnimationFrame(() => {
47 | const sz = this.computeSize();
48 | if (sz[0] < this.size[0]) sz[0] = this.size[0];
49 | if (sz[1] < this.size[1]) sz[1] = this.size[1];
50 | this.onResize?.(sz);
51 | app.graph.setDirtyCanvas(true, false);
52 | });
53 | } catch (error) {
54 | // Silent error handling
55 | }
56 | }
57 |
58 | const onExecuted = nodeType.prototype.onExecuted;
59 | nodeType.prototype.onExecuted = function (message) {
60 | try {
61 | onExecuted?.apply(this, arguments);
62 | if (message.enhanced_prompt) {
63 | populateEnhancedPrompt.call(this, message.enhanced_prompt);
64 | } else if (message.ui && message.ui.enhanced_prompt) {
65 | populateEnhancedPrompt.call(this, message.ui.enhanced_prompt);
66 | } else if (message.text) {
67 | populateEnhancedPrompt.call(this, message.text);
68 | }
69 | } catch (error) {
70 | // Silent error handling
71 | }
72 | };
73 | }
74 | },
75 | });
76 |
77 | window.DSD_ENHANCED_PROMPT_LOADED = true;
--------------------------------------------------------------------------------