├── .github └── workflows │ └── typos.yml ├── .gitignore ├── LICENSE.txt ├── README.md ├── _typos.toml ├── javascript └── additional_networks.js ├── models └── lora │ └── .keep ├── preload.py ├── scripts ├── additional_networks.py ├── addnet_xyz_grid_support.py ├── lora_compvis.py ├── metadata_editor.py ├── model_util.py ├── safetensors_hack.py └── util.py └── style.css /.github/workflows/typos.yml: -------------------------------------------------------------------------------- 1 | --- 2 | # yamllint disable rule:line-length 3 | name: Typos 4 | 5 | on: # yamllint disable-line rule:truthy 6 | push: 7 | pull_request: 8 | types: 9 | - opened 10 | - synchronize 11 | - reopened 12 | 13 | jobs: 14 | build: 15 | runs-on: ubuntu-latest 16 | 17 | steps: 18 | - uses: actions/checkout@v3 19 | 20 | - name: typos-action 21 | uses: crate-ci/typos@v1.13.10 22 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__ 2 | models/ 3 | /hashes.json 4 | -------------------------------------------------------------------------------- /LICENSE.txt: -------------------------------------------------------------------------------- 1 | GNU AFFERO GENERAL PUBLIC LICENSE 2 | Version 3, 19 November 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 Affero General Public License is a free, copyleft license for 11 | software and other kinds of works, specifically designed to ensure 12 | cooperation with the community in the case of network server software. 13 | 14 | The licenses for most software and other practical works are designed 15 | to take away your freedom to share and change the works. By contrast, 16 | our General Public Licenses are intended to guarantee your freedom to 17 | share and change all versions of a program--to make sure it remains free 18 | software for all its users. 19 | 20 | When we speak of free software, we are referring to freedom, not 21 | price. Our General Public Licenses are designed to make sure that you 22 | have the freedom to distribute copies of free software (and charge for 23 | them if you wish), that you receive source code or can get it if you 24 | want it, that you can change the software or use pieces of it in new 25 | free programs, and that you know you can do these things. 26 | 27 | Developers that use our General Public Licenses protect your rights 28 | with two steps: (1) assert copyright on the software, and (2) offer 29 | you this License which gives you legal permission to copy, distribute 30 | and/or modify the software. 31 | 32 | A secondary benefit of defending all users' freedom is that 33 | improvements made in alternate versions of the program, if they 34 | receive widespread use, become available for other developers to 35 | incorporate. Many developers of free software are heartened and 36 | encouraged by the resulting cooperation. However, in the case of 37 | software used on network servers, this result may fail to come about. 38 | The GNU General Public License permits making a modified version and 39 | letting the public access it on a server without ever releasing its 40 | source code to the public. 41 | 42 | The GNU Affero General Public License is designed specifically to 43 | ensure that, in such cases, the modified source code becomes available 44 | to the community. It requires the operator of a network server to 45 | provide the source code of the modified version running there to the 46 | users of that server. Therefore, public use of a modified version, on 47 | a publicly accessible server, gives the public access to the source 48 | code of the modified version. 49 | 50 | An older license, called the Affero General Public License and 51 | published by Affero, was designed to accomplish similar goals. This is 52 | a different license, not a version of the Affero GPL, but Affero has 53 | released a new version of the Affero GPL which permits relicensing under 54 | this license. 55 | 56 | The precise terms and conditions for copying, distribution and 57 | modification follow. 58 | 59 | TERMS AND CONDITIONS 60 | 61 | 0. Definitions. 62 | 63 | "This License" refers to version 3 of the GNU Affero General Public License. 64 | 65 | "Copyright" also means copyright-like laws that apply to other kinds of 66 | works, such as semiconductor masks. 67 | 68 | "The Program" refers to any copyrightable work licensed under this 69 | License. Each licensee is addressed as "you". "Licensees" and 70 | "recipients" may be individuals or organizations. 71 | 72 | To "modify" a work means to copy from or adapt all or part of the work 73 | in a fashion requiring copyright permission, other than the making of an 74 | exact copy. The resulting work is called a "modified version" of the 75 | earlier work or a work "based on" the earlier work. 76 | 77 | A "covered work" means either the unmodified Program or a work based 78 | on the Program. 79 | 80 | To "propagate" a work means to do anything with it that, without 81 | permission, would make you directly or secondarily liable for 82 | infringement under applicable copyright law, except executing it on a 83 | computer or modifying a private copy. Propagation includes copying, 84 | distribution (with or without modification), making available to the 85 | public, and in some countries other activities as well. 86 | 87 | To "convey" a work means any kind of propagation that enables other 88 | parties to make or receive copies. Mere interaction with a user through 89 | a computer network, with no transfer of a copy, is not conveying. 90 | 91 | An interactive user interface displays "Appropriate Legal Notices" 92 | to the extent that it includes a convenient and prominently visible 93 | feature that (1) displays an appropriate copyright notice, and (2) 94 | tells the user that there is no warranty for the work (except to the 95 | extent that warranties are provided), that licensees may convey the 96 | work under this License, and how to view a copy of this License. If 97 | the interface presents a list of user commands or options, such as a 98 | menu, a prominent item in the list meets this criterion. 99 | 100 | 1. Source Code. 101 | 102 | The "source code" for a work means the preferred form of the work 103 | for making modifications to it. "Object code" means any non-source 104 | form of a work. 105 | 106 | A "Standard Interface" means an interface that either is an official 107 | standard defined by a recognized standards body, or, in the case of 108 | interfaces specified for a particular programming language, one that 109 | is widely used among developers working in that language. 110 | 111 | The "System Libraries" of an executable work include anything, other 112 | than the work as a whole, that (a) is included in the normal form of 113 | packaging a Major Component, but which is not part of that Major 114 | Component, and (b) serves only to enable use of the work with that 115 | Major Component, or to implement a Standard Interface for which an 116 | implementation is available to the public in source code form. A 117 | "Major Component", in this context, means a major essential component 118 | (kernel, window system, and so on) of the specific operating system 119 | (if any) on which the executable work runs, or a compiler used to 120 | produce the work, or an object code interpreter used to run it. 121 | 122 | The "Corresponding Source" for a work in object code form means all 123 | the source code needed to generate, install, and (for an executable 124 | work) run the object code and to modify the work, including scripts to 125 | control those activities. However, it does not include the work's 126 | System Libraries, or general-purpose tools or generally available free 127 | programs which are used unmodified in performing those activities but 128 | which are not part of the work. For example, Corresponding Source 129 | includes interface definition files associated with source files for 130 | the work, and the source code for shared libraries and dynamically 131 | linked subprograms that the work is specifically designed to require, 132 | such as by intimate data communication or control flow between those 133 | subprograms and other parts of the work. 134 | 135 | The Corresponding Source need not include anything that users 136 | can regenerate automatically from other parts of the Corresponding 137 | Source. 138 | 139 | The Corresponding Source for a work in source code form is that 140 | same work. 141 | 142 | 2. Basic Permissions. 143 | 144 | All rights granted under this License are granted for the term of 145 | copyright on the Program, and are irrevocable provided the stated 146 | conditions are met. This License explicitly affirms your unlimited 147 | permission to run the unmodified Program. The output from running a 148 | covered work is covered by this License only if the output, given its 149 | content, constitutes a covered work. This License acknowledges your 150 | rights of fair use or other equivalent, as provided by copyright law. 151 | 152 | You may make, run and propagate covered works that you do not 153 | convey, without conditions so long as your license otherwise remains 154 | in force. You may convey covered works to others for the sole purpose 155 | of having them make modifications exclusively for you, or provide you 156 | with facilities for running those works, provided that you comply with 157 | the terms of this License in conveying all material for which you do 158 | not control copyright. Those thus making or running the covered works 159 | for you must do so exclusively on your behalf, under your direction 160 | and control, on terms that prohibit them from making any copies of 161 | your copyrighted material outside their relationship with you. 162 | 163 | Conveying under any other circumstances is permitted solely under 164 | the conditions stated below. Sublicensing is not allowed; section 10 165 | makes it unnecessary. 166 | 167 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 168 | 169 | No covered work shall be deemed part of an effective technological 170 | measure under any applicable law fulfilling obligations under article 171 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 172 | similar laws prohibiting or restricting circumvention of such 173 | measures. 174 | 175 | When you convey a covered work, you waive any legal power to forbid 176 | circumvention of technological measures to the extent such circumvention 177 | is effected by exercising rights under this License with respect to 178 | the covered work, and you disclaim any intention to limit operation or 179 | modification of the work as a means of enforcing, against the work's 180 | users, your or third parties' legal rights to forbid circumvention of 181 | technological measures. 182 | 183 | 4. Conveying Verbatim Copies. 184 | 185 | You may convey verbatim copies of the Program's source code as you 186 | receive it, in any medium, provided that you conspicuously and 187 | appropriately publish on each copy an appropriate copyright notice; 188 | keep intact all notices stating that this License and any 189 | non-permissive terms added in accord with section 7 apply to the code; 190 | keep intact all notices of the absence of any warranty; and give all 191 | recipients a copy of this License along with the Program. 192 | 193 | You may charge any price or no price for each copy that you convey, 194 | and you may offer support or warranty protection for a fee. 195 | 196 | 5. Conveying Modified Source Versions. 197 | 198 | You may convey a work based on the Program, or the modifications to 199 | produce it from the Program, in the form of source code under the 200 | terms of section 4, provided that you also meet all of these conditions: 201 | 202 | a) The work must carry prominent notices stating that you modified 203 | it, and giving a relevant date. 204 | 205 | b) The work must carry prominent notices stating that it is 206 | released under this License and any conditions added under section 207 | 7. This requirement modifies the requirement in section 4 to 208 | "keep intact all notices". 209 | 210 | c) You must license the entire work, as a whole, under this 211 | License to anyone who comes into possession of a copy. This 212 | License will therefore apply, along with any applicable section 7 213 | additional terms, to the whole of the work, and all its parts, 214 | regardless of how they are packaged. This License gives no 215 | permission to license the work in any other way, but it does not 216 | invalidate such permission if you have separately received it. 217 | 218 | d) If the work has interactive user interfaces, each must display 219 | Appropriate Legal Notices; however, if the Program has interactive 220 | interfaces that do not display Appropriate Legal Notices, your 221 | work need not make them do so. 222 | 223 | A compilation of a covered work with other separate and independent 224 | works, which are not by their nature extensions of the covered work, 225 | and which are not combined with it such as to form a larger program, 226 | in or on a volume of a storage or distribution medium, is called an 227 | "aggregate" if the compilation and its resulting copyright are not 228 | used to limit the access or legal rights of the compilation's users 229 | beyond what the individual works permit. Inclusion of a covered work 230 | in an aggregate does not cause this License to apply to the other 231 | parts of the aggregate. 232 | 233 | 6. Conveying Non-Source Forms. 234 | 235 | You may convey a covered work in object code form under the terms 236 | of sections 4 and 5, provided that you also convey the 237 | machine-readable Corresponding Source under the terms of this License, 238 | in one of these ways: 239 | 240 | a) Convey the object code in, or embodied in, a physical product 241 | (including a physical distribution medium), accompanied by the 242 | Corresponding Source fixed on a durable physical medium 243 | customarily used for software interchange. 244 | 245 | b) Convey the object code in, or embodied in, a physical product 246 | (including a physical distribution medium), accompanied by a 247 | written offer, valid for at least three years and valid for as 248 | long as you offer spare parts or customer support for that product 249 | model, to give anyone who possesses the object code either (1) a 250 | copy of the Corresponding Source for all the software in the 251 | product that is covered by this License, on a durable physical 252 | medium customarily used for software interchange, for a price no 253 | more than your reasonable cost of physically performing this 254 | conveying of source, or (2) access to copy the 255 | Corresponding Source from a network server at no charge. 256 | 257 | c) Convey individual copies of the object code with a copy of the 258 | written offer to provide the Corresponding Source. This 259 | alternative is allowed only occasionally and noncommercially, and 260 | only if you received the object code with such an offer, in accord 261 | with subsection 6b. 262 | 263 | d) Convey the object code by offering access from a designated 264 | place (gratis or for a charge), and offer equivalent access to the 265 | Corresponding Source in the same way through the same place at no 266 | further charge. You need not require recipients to copy the 267 | Corresponding Source along with the object code. If the place to 268 | copy the object code is a network server, the Corresponding Source 269 | may be on a different server (operated by you or a third party) 270 | that supports equivalent copying facilities, provided you maintain 271 | clear directions next to the object code saying where to find the 272 | Corresponding Source. Regardless of what server hosts the 273 | Corresponding Source, you remain obligated to ensure that it is 274 | available for as long as needed to satisfy these requirements. 275 | 276 | e) Convey the object code using peer-to-peer transmission, provided 277 | you inform other peers where the object code and Corresponding 278 | Source of the work are being offered to the general public at no 279 | charge under subsection 6d. 280 | 281 | A separable portion of the object code, whose source code is excluded 282 | from the Corresponding Source as a System Library, need not be 283 | included in conveying the object code work. 284 | 285 | A "User Product" is either (1) a "consumer product", which means any 286 | tangible personal property which is normally used for personal, family, 287 | or household purposes, or (2) anything designed or sold for incorporation 288 | into a dwelling. In determining whether a product is a consumer product, 289 | doubtful cases shall be resolved in favor of coverage. For a particular 290 | product received by a particular user, "normally used" refers to a 291 | typical or common use of that class of product, regardless of the status 292 | of the particular user or of the way in which the particular user 293 | actually uses, or expects or is expected to use, the product. A product 294 | is a consumer product regardless of whether the product has substantial 295 | commercial, industrial or non-consumer uses, unless such uses represent 296 | the only significant mode of use of the product. 297 | 298 | "Installation Information" for a User Product means any methods, 299 | procedures, authorization keys, or other information required to install 300 | and execute modified versions of a covered work in that User Product from 301 | a modified version of its Corresponding Source. The information must 302 | suffice to ensure that the continued functioning of the modified object 303 | code is in no case prevented or interfered with solely because 304 | modification has been made. 305 | 306 | If you convey an object code work under this section in, or with, or 307 | specifically for use in, a User Product, and the conveying occurs as 308 | part of a transaction in which the right of possession and use of the 309 | User Product is transferred to the recipient in perpetuity or for a 310 | fixed term (regardless of how the transaction is characterized), the 311 | Corresponding Source conveyed under this section must be accompanied 312 | by the Installation Information. But this requirement does not apply 313 | if neither you nor any third party retains the ability to install 314 | modified object code on the User Product (for example, the work has 315 | been installed in ROM). 316 | 317 | The requirement to provide Installation Information does not include a 318 | requirement to continue to provide support service, warranty, or updates 319 | for a work that has been modified or installed by the recipient, or for 320 | the User Product in which it has been modified or installed. Access to a 321 | network may be denied when the modification itself materially and 322 | adversely affects the operation of the network or violates the rules and 323 | protocols for communication across the network. 324 | 325 | Corresponding Source conveyed, and Installation Information provided, 326 | in accord with this section must be in a format that is publicly 327 | documented (and with an implementation available to the public in 328 | source code form), and must require no special password or key for 329 | unpacking, reading or copying. 330 | 331 | 7. Additional Terms. 332 | 333 | "Additional permissions" are terms that supplement the terms of this 334 | License by making exceptions from one or more of its conditions. 335 | Additional permissions that are applicable to the entire Program shall 336 | be treated as though they were included in this License, to the extent 337 | that they are valid under applicable law. If additional permissions 338 | apply only to part of the Program, that part may be used separately 339 | under those permissions, but the entire Program remains governed by 340 | this License without regard to the additional permissions. 341 | 342 | When you convey a copy of a covered work, you may at your option 343 | remove any additional permissions from that copy, or from any part of 344 | it. (Additional permissions may be written to require their own 345 | removal in certain cases when you modify the work.) You may place 346 | additional permissions on material, added by you to a covered work, 347 | for which you have or can give appropriate copyright permission. 348 | 349 | Notwithstanding any other provision of this License, for material you 350 | add to a covered work, you may (if authorized by the copyright holders of 351 | that material) supplement the terms of this License with terms: 352 | 353 | a) Disclaiming warranty or limiting liability differently from the 354 | terms of sections 15 and 16 of this License; or 355 | 356 | b) Requiring preservation of specified reasonable legal notices or 357 | author attributions in that material or in the Appropriate Legal 358 | Notices displayed by works containing it; or 359 | 360 | c) Prohibiting misrepresentation of the origin of that material, or 361 | requiring that modified versions of such material be marked in 362 | reasonable ways as different from the original version; or 363 | 364 | d) Limiting the use for publicity purposes of names of licensors or 365 | authors of the material; or 366 | 367 | e) Declining to grant rights under trademark law for use of some 368 | trade names, trademarks, or service marks; or 369 | 370 | f) Requiring indemnification of licensors and authors of that 371 | material by anyone who conveys the material (or modified versions of 372 | it) with contractual assumptions of liability to the recipient, for 373 | any liability that these contractual assumptions directly impose on 374 | those licensors and authors. 375 | 376 | All other non-permissive additional terms are considered "further 377 | restrictions" within the meaning of section 10. If the Program as you 378 | received it, or any part of it, contains a notice stating that it is 379 | governed by this License along with a term that is a further 380 | restriction, you may remove that term. If a license document contains 381 | a further restriction but permits relicensing or conveying under this 382 | License, you may add to a covered work material governed by the terms 383 | of that license document, provided that the further restriction does 384 | not survive such relicensing or conveying. 385 | 386 | If you add terms to a covered work in accord with this section, you 387 | must place, in the relevant source files, a statement of the 388 | additional terms that apply to those files, or a notice indicating 389 | where to find the applicable terms. 390 | 391 | Additional terms, permissive or non-permissive, may be stated in the 392 | form of a separately written license, or stated as exceptions; 393 | the above requirements apply either way. 394 | 395 | 8. Termination. 396 | 397 | You may not propagate or modify a covered work except as expressly 398 | provided under this License. Any attempt otherwise to propagate or 399 | modify it is void, and will automatically terminate your rights under 400 | this License (including any patent licenses granted under the third 401 | paragraph of section 11). 402 | 403 | However, if you cease all violation of this License, then your 404 | license from a particular copyright holder is reinstated (a) 405 | provisionally, unless and until the copyright holder explicitly and 406 | finally terminates your license, and (b) permanently, if the copyright 407 | holder fails to notify you of the violation by some reasonable means 408 | prior to 60 days after the cessation. 409 | 410 | Moreover, your license from a particular copyright holder is 411 | reinstated permanently if the copyright holder notifies you of the 412 | violation by some reasonable means, this is the first time you have 413 | received notice of violation of this License (for any work) from that 414 | copyright holder, and you cure the violation prior to 30 days after 415 | your receipt of the notice. 416 | 417 | Termination of your rights under this section does not terminate the 418 | licenses of parties who have received copies or rights from you under 419 | this License. If your rights have been terminated and not permanently 420 | reinstated, you do not qualify to receive new licenses for the same 421 | material under section 10. 422 | 423 | 9. Acceptance Not Required for Having Copies. 424 | 425 | You are not required to accept this License in order to receive or 426 | run a copy of the Program. Ancillary propagation of a covered work 427 | occurring solely as a consequence of using peer-to-peer transmission 428 | to receive a copy likewise does not require acceptance. However, 429 | nothing other than this License grants you permission to propagate or 430 | modify any covered work. These actions infringe copyright if you do 431 | not accept this License. Therefore, by modifying or propagating a 432 | covered work, you indicate your acceptance of this License to do so. 433 | 434 | 10. Automatic Licensing of Downstream Recipients. 435 | 436 | Each time you convey a covered work, the recipient automatically 437 | receives a license from the original licensors, to run, modify and 438 | propagate that work, subject to this License. You are not responsible 439 | for enforcing compliance by third parties with this License. 440 | 441 | An "entity transaction" is a transaction transferring control of an 442 | organization, or substantially all assets of one, or subdividing an 443 | organization, or merging organizations. If propagation of a covered 444 | work results from an entity transaction, each party to that 445 | transaction who receives a copy of the work also receives whatever 446 | licenses to the work the party's predecessor in interest had or could 447 | give under the previous paragraph, plus a right to possession of the 448 | Corresponding Source of the work from the predecessor in interest, if 449 | the predecessor has it or can get it with reasonable efforts. 450 | 451 | You may not impose any further restrictions on the exercise of the 452 | rights granted or affirmed under this License. For example, you may 453 | not impose a license fee, royalty, or other charge for exercise of 454 | rights granted under this License, and you may not initiate litigation 455 | (including a cross-claim or counterclaim in a lawsuit) alleging that 456 | any patent claim is infringed by making, using, selling, offering for 457 | sale, or importing the Program or any portion of it. 458 | 459 | 11. Patents. 460 | 461 | A "contributor" is a copyright holder who authorizes use under this 462 | License of the Program or a work on which the Program is based. The 463 | work thus licensed is called the contributor's "contributor version". 464 | 465 | A contributor's "essential patent claims" are all patent claims 466 | owned or controlled by the contributor, whether already acquired or 467 | hereafter acquired, that would be infringed by some manner, permitted 468 | by this License, of making, using, or selling its contributor version, 469 | but do not include claims that would be infringed only as a 470 | consequence of further modification of the contributor version. For 471 | purposes of this definition, "control" includes the right to grant 472 | patent sublicenses in a manner consistent with the requirements of 473 | this License. 474 | 475 | Each contributor grants you a non-exclusive, worldwide, royalty-free 476 | patent license under the contributor's essential patent claims, to 477 | make, use, sell, offer for sale, import and otherwise run, modify and 478 | propagate the contents of its contributor version. 479 | 480 | In the following three paragraphs, a "patent license" is any express 481 | agreement or commitment, however denominated, not to enforce a patent 482 | (such as an express permission to practice a patent or covenant not to 483 | sue for patent infringement). To "grant" such a patent license to a 484 | party means to make such an agreement or commitment not to enforce a 485 | patent against the party. 486 | 487 | If you convey a covered work, knowingly relying on a patent license, 488 | and the Corresponding Source of the work is not available for anyone 489 | to copy, free of charge and under the terms of this License, through a 490 | publicly available network server or other readily accessible means, 491 | then you must either (1) cause the Corresponding Source to be so 492 | available, or (2) arrange to deprive yourself of the benefit of the 493 | patent license for this particular work, or (3) arrange, in a manner 494 | consistent with the requirements of this License, to extend the patent 495 | license to downstream recipients. "Knowingly relying" means you have 496 | actual knowledge that, but for the patent license, your conveying the 497 | covered work in a country, or your recipient's use of the covered work 498 | in a country, would infringe one or more identifiable patents in that 499 | country that you have reason to believe are valid. 500 | 501 | If, pursuant to or in connection with a single transaction or 502 | arrangement, you convey, or propagate by procuring conveyance of, a 503 | covered work, and grant a patent license to some of the parties 504 | receiving the covered work authorizing them to use, propagate, modify 505 | or convey a specific copy of the covered work, then the patent license 506 | you grant is automatically extended to all recipients of the covered 507 | work and works based on it. 508 | 509 | A patent license is "discriminatory" if it does not include within 510 | the scope of its coverage, prohibits the exercise of, or is 511 | conditioned on the non-exercise of one or more of the rights that are 512 | specifically granted under this License. You may not convey a covered 513 | work if you are a party to an arrangement with a third party that is 514 | in the business of distributing software, under which you make payment 515 | to the third party based on the extent of your activity of conveying 516 | the work, and under which the third party grants, to any of the 517 | parties who would receive the covered work from you, a discriminatory 518 | patent license (a) in connection with copies of the covered work 519 | conveyed by you (or copies made from those copies), or (b) primarily 520 | for and in connection with specific products or compilations that 521 | contain the covered work, unless you entered into that arrangement, 522 | or that patent license was granted, prior to 28 March 2007. 523 | 524 | Nothing in this License shall be construed as excluding or limiting 525 | any implied license or other defenses to infringement that may 526 | otherwise be available to you under applicable patent law. 527 | 528 | 12. No Surrender of Others' Freedom. 529 | 530 | If conditions are imposed on you (whether by court order, agreement or 531 | otherwise) that contradict the conditions of this License, they do not 532 | excuse you from the conditions of this License. If you cannot convey a 533 | covered work so as to satisfy simultaneously your obligations under this 534 | License and any other pertinent obligations, then as a consequence you may 535 | not convey it at all. For example, if you agree to terms that obligate you 536 | to collect a royalty for further conveying from those to whom you convey 537 | the Program, the only way you could satisfy both those terms and this 538 | License would be to refrain entirely from conveying the Program. 539 | 540 | 13. Remote Network Interaction; Use with the GNU General Public License. 541 | 542 | Notwithstanding any other provision of this License, if you modify the 543 | Program, your modified version must prominently offer all users 544 | interacting with it remotely through a computer network (if your version 545 | supports such interaction) an opportunity to receive the Corresponding 546 | Source of your version by providing access to the Corresponding Source 547 | from a network server at no charge, through some standard or customary 548 | means of facilitating copying of software. This Corresponding Source 549 | shall include the Corresponding Source for any work covered by version 3 550 | of the GNU General Public License that is incorporated pursuant to the 551 | following paragraph. 552 | 553 | Notwithstanding any other provision of this License, you have 554 | permission to link or combine any covered work with a work licensed 555 | under version 3 of the GNU General Public License into a single 556 | combined work, and to convey the resulting work. The terms of this 557 | License will continue to apply to the part which is the covered work, 558 | but the work with which it is combined will remain governed by version 559 | 3 of the GNU General Public License. 560 | 561 | 14. Revised Versions of this License. 562 | 563 | The Free Software Foundation may publish revised and/or new versions of 564 | the GNU Affero General Public License from time to time. Such new versions 565 | will be similar in spirit to the present version, but may differ in detail to 566 | address new problems or concerns. 567 | 568 | Each version is given a distinguishing version number. If the 569 | Program specifies that a certain numbered version of the GNU Affero General 570 | Public License "or any later version" applies to it, you have the 571 | option of following the terms and conditions either of that numbered 572 | version or of any later version published by the Free Software 573 | Foundation. If the Program does not specify a version number of the 574 | GNU Affero General Public License, you may choose any version ever published 575 | by the Free Software Foundation. 576 | 577 | If the Program specifies that a proxy can decide which future 578 | versions of the GNU Affero General Public License can be used, that proxy's 579 | public statement of acceptance of a version permanently authorizes you 580 | to choose that version for the Program. 581 | 582 | Later license versions may give you additional or different 583 | permissions. However, no additional obligations are imposed on any 584 | author or copyright holder as a result of your choosing to follow a 585 | later version. 586 | 587 | 15. Disclaimer of Warranty. 588 | 589 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 590 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 591 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 592 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 593 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 594 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 595 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 596 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 597 | 598 | 16. Limitation of Liability. 599 | 600 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 601 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 602 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 603 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 604 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 605 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 606 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 607 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 608 | SUCH DAMAGES. 609 | 610 | 17. Interpretation of Sections 15 and 16. 611 | 612 | If the disclaimer of warranty and limitation of liability provided 613 | above cannot be given local legal effect according to their terms, 614 | reviewing courts shall apply local law that most closely approximates 615 | an absolute waiver of all civil liability in connection with the 616 | Program, unless a warranty or assumption of liability accompanies a 617 | copy of the Program in return for a fee. 618 | 619 | END OF TERMS AND CONDITIONS 620 | 621 | How to Apply These Terms to Your New Programs 622 | 623 | If you develop a new program, and you want it to be of the greatest 624 | possible use to the public, the best way to achieve this is to make it 625 | free software which everyone can redistribute and change under these terms. 626 | 627 | To do so, attach the following notices to the program. It is safest 628 | to attach them to the start of each source file to most effectively 629 | state the exclusion of warranty; and each file should have at least 630 | the "copyright" line and a pointer to where the full notice is found. 631 | 632 | 633 | Copyright (C) 634 | 635 | This program is free software: you can redistribute it and/or modify 636 | it under the terms of the GNU Affero General Public License as published by 637 | the Free Software Foundation, either version 3 of the License, or 638 | (at your option) any later version. 639 | 640 | This program is distributed in the hope that it will be useful, 641 | but WITHOUT ANY WARRANTY; without even the implied warranty of 642 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 643 | GNU Affero General Public License for more details. 644 | 645 | You should have received a copy of the GNU Affero General Public License 646 | along with this program. If not, see . 647 | 648 | Also add information on how to contact you by electronic and paper mail. 649 | 650 | If your software can interact with users remotely through a computer 651 | network, you should also make sure that it provides a way for users to 652 | get its source. For example, if your program is a web application, its 653 | interface could display a "Source" link that leads users to an archive 654 | of the code. There are many ways you could offer source, and different 655 | solutions will be better for different programs; see section 13 for the 656 | specific requirements. 657 | 658 | You should also get your employer (if you work as a programmer) or school, 659 | if any, to sign a "copyright disclaimer" for the program, if necessary. 660 | For more information on this, and how to apply and follow the GNU AGPL, see 661 | . 662 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## Additional Networks for generating images 2 | 3 | 日本語の文章は下のほうにあります。 4 | 5 | [__Change History__](#change-history) is moved to the bottom of the page. 6 | 更新履歴は[ページ末尾](#change-history)に移しました。 7 | 8 | Stable Diffusion web UI now seems to support LoRA trained by ``sd-scripts`` Thank you for great work!!! 9 | 10 | 11 | ## About 12 | 13 | This extension is for [AUTOMATIC1111's Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui), allows the Web UI to add some networks (e.g. LoRA) to the original Stable Diffusion model to generate images. Currently LoRA is supported. The addition is on-the-fly, the merging is not required. 14 | 15 | This extension supports the LoRA models (*.ckpt or *.safetensors) trained by our scripts in [sd-scripts](https://github.com/kohya-ss/sd-scripts). The models from other LoRA implementations are not supported. 16 | 17 | This extension does not support training. 18 | 19 | Other networks other than LoRA may be supported in the future. 20 | 21 | ## Installation 22 | 23 | 1. Open "Extensions" tab. 24 | 1. Open "Install from URL" tab in the tab. 25 | 1. Enter URL of this repo to "URL for extension's git repository". 26 | 1. Press "Install" button. 27 | 1. Restart Web UI. 28 | 29 | ## How to use 30 | 31 | Put the LoRA models (`*.pt`, `*.ckpt` or `*.safetensors`) inside the `sd-webui-additional-networks/models/LoRA` folder. 32 | 33 | Open __"Additional Networks"__ panel from the left bottom of Web UI. 34 | 35 | Press __"Refresh models"__ to update the models list. 36 | 37 | Select __"LoRA"__ for __"Network module 1"__. 38 | 39 | Choose __the name of the LoRA model file__ in __"Model 1"__. 40 | 41 | Set __the weight__ of the model (negative weight might be working but unexpected.) 42 | 43 | Repeat them for the module/model/weight 2 to 5 if you have other models. Models are applied in the order of 1 to 5. 44 | 45 | You can generate images with the model with these additional networks. 46 | 47 | ## X/Y plot 48 | 49 | If you use LoRA models to plot, put the comma separated list of the model names into ``AddNet Model X`` 50 | 51 | ![image](https://user-images.githubusercontent.com/52813779/212444037-8ccd9157-c341-4eb4-82b4-64e3c8ee0237.png) 52 | 53 | You can get the list of models with the button next to ``Values``. Please select any model in ``Model ?`` at ``Additional Networks`` in order to make the button work. Models in the same folder as the model will be listed. 54 | 55 | ![image](https://user-images.githubusercontent.com/52813779/212443639-97779d8d-0f7e-47f0-919c-41f053fe28a9.png) 56 | 57 | The metadata of the model can be drawn as legends. Move to ``Settings`` tab, select ``Additional Networks`` at left bottom, and set ``Metadata to show``. Available values are in ``Network metadata`` textbox in ``Additional Networks`` tab. 58 | 59 | ![image](https://user-images.githubusercontent.com/52813779/212443781-1f4c442e-c2f3-47f8-b698-25fbe981f633.png) 60 | 61 | ## Specify target region of LoRA by mask (__experimental__) 62 | 63 | Open `Extra args` and drop a mask image to `mask image`. 64 | 65 | By specifying with the mask image, each LoRA model can be applied only to the specified region of the image. Currently, only three models (Models 1 to 3) can be masked. 66 | 67 | The mask image is RGB image, with each channel (R, G and B) corresponding to LoRA models 1 to 3. Each channel can be overlapped. For example, yellow area (R and G) is applied to LoRA model 1 and 2. The range of values is 0 to 255, corresponding to a LoRA weight of 0 to 1. 68 | 69 | It can be combined with ControlNet. 70 | 71 | | |without ControlNet|with ControlNet| 72 | |:----:|:----:|:----:| 73 | |no LoRA||| 74 | |with LoRA, no mask||| 75 | |with Lora, with mask||| 76 | | |pose|mask| 77 | | || 78 | 79 | Sample images are generated with [wd-1-5-beta2-aesthetic-fp16.safetensors](https://huggingface.co/waifu-diffusion/wd-1-5-beta2) and three LoRAs: two character LoRAs (model 1 and 2, masked, weight=1.0) and one style LoRA (model 4, not masked, weight=0.8). Used ControlNet is [diff_control_wd15beta2_pose.safetensors](https://huggingface.co/furusu/ControlNet). 80 | 81 | ### Difference from 'Latent Couple extension' and 'Composable LoRA' 82 | 83 | 'Latent Couple extension' masks the output of U-Net for each sub-prompt (AND-separated prompts), while our implementation masks the output of LoRA at each layer of U-Net. The mask is resized according to the tensor shape of each layer, so the resolution is particularly coarse at the deeper layers. 84 | 85 | 'Composable LoRA' controls the area via 'Latent Couple extension' by switching LoRA on or off for each sub-prompt, but this implementation works alone. 86 | 87 | This implementation does not work for all modules in LoRA (the modules associated with Text Encoder are not masked), and due to the coarse resolution, it is not possible to completely separate areas. 88 | 89 | ## この Web UI 拡張について 90 | 91 | LoRA などのネットワークを元の Stable Diffusion に追加し、画像生成を行うための拡張です。現在は LoRA のみ対応しています。 92 | 93 | この拡張で使えるのは[sd-scripts](https://github.com/kohya-ss/sd-scripts)リポジトリで学習した LoRA のモデル（\*.ckpt または \*.safetensors）です。他の LoRA リポジトリで学習したモデルは対応していません。 94 | 95 | この拡張単体では学習はできません。 96 | 97 | 将来的に LoRA 以外のネットワークについてもサポートするかもしれません。 98 | 99 | ## インストール 100 | 101 | 1. Web UI で "Extensions" タブを開きます。 102 | 1. さらに "Install from URL" タブを開きます。 103 | 1. "URL for extension's git repository" 欄にこのリポジトリの URL を入れます。 104 | 1. "Install"ボタンを押してインストールします。 105 | 1. Web UI を再起動してください。 106 | 107 | ## 使用法 108 | 109 | 学習した LoRA のモデル(`*.pt`, `*.ckpt`, `*.safetensors`)を`sd-webui-additional-networks/models/LoRA`に置きます。 110 | 111 | Web UI の左下のほうの __"Additional Networks"__ のパネルを開きます。 112 | 113 | __"Network module 1"__ で __"LoRA"__ を選択してください。 114 | 115 | __"Refresh models"__ で LoRA モデルのリストを更新します。 116 | 117 | __"Model 1"__ に学習した LoRA のモデル名を選択します。 118 | 119 | __"Weight"__ にこのモデルの __重み__ を指定します（負の値も指定できますがどんな効果があるかは未知数です）。 120 | 121 | 追加のモデルがある場合は 2～5 に指定してください。モデルは 1~5 の順番で適用されます。 122 | 123 | 以上を指定すると、それぞれのモデルが適用された状態で画像生成されます。 124 | 125 | ## X/Y plot 126 | 127 | LoRAモデルをX/Y plotの値（選択対象）として使う場合は、カンマ区切りのモデルのリストを与える必要があります。 128 | 129 | ![image](https://user-images.githubusercontent.com/52813779/212444037-8ccd9157-c341-4eb4-82b4-64e3c8ee0237.png) 130 | 131 | モデルのリストは選択肢の隣にあるボタンで取得できます。いずれかのモデルを ``Additional Networks`` の ``Model ?`` で選択しておいてください。そのモデルと同じフォルダにあるモデルの一覧が取得されます。 132 | 133 | ![image](https://user-images.githubusercontent.com/52813779/212443639-97779d8d-0f7e-47f0-919c-41f053fe28a9.png) 134 | 135 | モデルのメタデータ（学習時のパラメータなど）をX/Y plotのラベルに使用できます。Web UI上部の ``Settings`` タブを開き、左下から ``Additional Networks`` を選び、 ``Metadata to show`` にカンマ区切りで項目名を指定してください（``ss_learning_rate, ss_num_epochs`` のような感じになります）。使える値は ``Additional Networks`` の ``Network metadata`` 欄にある値です。 136 | 137 | ![image](https://user-images.githubusercontent.com/52813779/212443781-1f4c442e-c2f3-47f8-b698-25fbe981f633.png) 138 | 139 | ## LoRA の領域別適用 __（実験的機能）__ 140 | 141 | 適用する領域をマスク画像で指定することで、それぞれの LoRA モデルを画像の指定した部分にのみ適用することができます。現在はモデル1~3の3つのみ領域指定可能です。 142 | 143 | マスク画像はカラーの画像で、RGBの各チャネルが LoRA モデル1から3に対応します。RGBの各チャネルは重ねることが可能です。たとえば黄色（RとGチャネル）の領域は、モデル1と2が有効になります。ピクセル値0から255がLoRAの適用率0から1に対応します（127なら重み0.5で適用するのと同じになります）。 144 | 145 | マスク画像は生成画像サイズにリサイズされて適用されます。 146 | 147 | ControlNetと組み合わせることも可能です（細かい位置指定にはControlNetとの組み合わせを推奨します）。 148 | 149 | 上のサンプルをご参照ください。 150 | 151 | ### Latent Couple extension、Composable LoRAとの違い 152 | 153 | Latent Couple extension はサブプロンプト（ANDで区切られたプロンプト）ごとに、U-Net の出力をマスクしますが、当実装では U-Net の各層で LoRA の出力をマスクします。マスクは各層のテンソル形状に応じてリサイズされるため、深い層では特に解像度が粗くなります。 154 | 155 | Composable LoRA はサブプロンプトごとに LoRA の適用有無を切り替えることで Latent Couple extension を経由して影響範囲を制御しますが、当実装では単独で動作します。 156 | 157 | 当実装はすべての LoRA モジュールに作用するわけではなく（Text Encoder に関連する LoRA モジュールはマスクされません）、また解像度が粗いため、完全に領域を分離することはできません。 158 | 159 | ## Change History 160 | 161 | - 23 May 2023, 2023/5/23 162 | - Fix an issue where the value of the `Weight` slider is not applied correctly. 163 | - `Weight`のスライダーの値が正しく反映されない場合がある不具合への対応を行いました。 164 | 165 | - 8 May 2023, 2023/5/8 166 | - Fix an issue where the models are not loaded correctly in the `Additional Networks` tab. 167 | - Fix an issue where `None` cannot be selected as a model in X/Y/Z plot. 168 | - `Additional Networks`タブでモデルが正しく読み込まれない不具合を修正しました。 169 | - X/Y/Z plotでモデルに `None` が選択できない不具合を修正しました。 170 | 171 | - 3 May 2023, 2023/5/3 172 | - Fix an issue where an error occurs when selecting a model in X/Y/Z plot. 173 | - X/Y/Z plotでモデル選択時にエラーとなる不具合を修正しました。 174 | - 6 Apr. 2023, 2023/4/6 175 | - Fix an issue where the `Hires. fix` does not work with mask. 176 | - 領域別LoRAでHires. fixが動作しない不具合を修正しました。 177 | - 30 Mar. 2023, 2023/3/30 178 | - Fix an issue where the `Save Metadata` button in the metadata editor does not work even if `Editing Enabled` is checked. 179 | - メタデータエディタで `Save Metadata` ボタンが `Editing Enabled` をチェックしても有効にならない不具合を修正しました。 180 | - 28 Mar. 2023, 2023/3/28 181 | - Fix style for Gradio 3.22. Thanks to space-nuko! 182 | - Please update Web UI to the latest version. 183 | - Gradio 3.22 のスタイルに対応しました。space-nuko氏に感謝します。 184 | - Web UIを最新版に更新願います。 185 | - 11 Mar. 2023, 2023/3/11 186 | - Leading spaces in each path in `Extra paths to scan for LoRA models` settings are ignored. Thanks to tsukimiya! 187 | - 設定の `Extra paths to scan for LoRA models` の各ディレクトリ名の先頭スペースを無視するよう変更しました。tsukimiya氏に感謝します。 188 | - 9 Mar. 2023, 2023/3/9: Release v0.5.1 189 | - Fix the model saved with `bf16` causes an error. https://github.com/kohya-ss/sd-webui-additional-networks/issues/127 190 | - Fix some Conv2d-3x3 LoRA modules are not effective. https://github.com/kohya-ss/sd-scripts/issues/275 191 | - Fix LoRA modules with higher dim (rank) > 320 causes an error. 192 | - `bf16` で学習されたモデルが読み込めない不具合を修正しました。 https://github.com/kohya-ss/sd-webui-additional-networks/issues/127 193 | - いくつかの Conv2d-3x3 LoRA モジュールが有効にならない不具合を修正しました。 https://github.com/kohya-ss/sd-scripts/issues/275 194 | - dim (rank) が 320 を超えるLoRAモデルが読み込めない不具合を修正しました。 195 | - 8 Mar. 2023, 2023/3/8: Release v0.5.0 196 | - Support current version of [LoCon](https://github.com/KohakuBlueleaf/LoCon). __Thank you very much KohakuBlueleaf for your help!__ 197 | - LoCon will be enhanced in the future. Compatibility for future versions is not guaranteed. 198 | - Support dynamic LoRA: different dimensions (ranks) and alpha for each module. 199 | - Support LoRA for Conv2d (extended to conv2d with a kernel size not 1x1). 200 | - Add masked LoRA feature (experimental.) 201 | - 現在のバージョンの [LoCon](https://github.com/KohakuBlueleaf/LoCon) をサポートしました。 KohakuBlueleaf 氏のご支援に深く感謝します。 202 | - LoCon が将来的に拡張された場合、それらのバージョンでの互換性は保証できません。 203 | - dynamic LoRA の機能を追加しました。各モジュールで異なる dimension (rank) や alpha を持つ LoRA が使えます。 204 | - Conv2d 拡張 LoRA をサポートしました。カーネルサイズが1x1でない Conv2d を対象とした LoRA が使えます。 205 | - LoRA の適用領域指定機能を追加しました（実験的機能）。 206 | 207 | 208 | Please read [Releases](https://github.com/kohya-ss/sd-webui-additional-networks/releases) for recent updates. 209 | 最近の更新情報は [Release](https://github.com/kohya-ss/sd-webui-additional-networks/releases) をご覧ください。 210 | 211 | -------------------------------------------------------------------------------- /_typos.toml: -------------------------------------------------------------------------------- 1 | # Files for typos 2 | # Instruction: https://github.com/marketplace/actions/typos-action#getting-started 3 | 4 | [default.extend-identifiers] 5 | 6 | [default.extend-words] 7 | 8 | [files] 9 | extend-exclude = ["_typos.toml"] 10 | -------------------------------------------------------------------------------- /javascript/additional_networks.js: -------------------------------------------------------------------------------- 1 | function addnet_switch_to_txt2img(){ 2 | switch_to_txt2img(); 3 | setTimeout(function() { gradioApp().getElementById("additional_networks_txt2img").scrollIntoView(); }, 100); 4 | return args_to_array(arguments); 5 | } 6 | 7 | function addnet_switch_to_img2img(){ 8 | switch_to_img2img(); 9 | setTimeout(function() { gradioApp().getElementById("additional_networks_img2img").scrollIntoView(); }, 100); 10 | return args_to_array(arguments); 11 | } 12 | 13 | function addnet_switch_to_metadata_editor(){ 14 | Array.from(gradioApp().querySelector('#tabs').querySelectorAll('button')).filter(e => e.textContent.trim() === "Additional Networks")[0].click(); 15 | return args_to_array(arguments); 16 | } 17 | 18 | function addnet_send_to_metadata_editor() { 19 | var module = arguments[0]; 20 | var model_path = arguments[1]; 21 | 22 | if (model_path == "None") { 23 | return args_to_array(arguments); 24 | } 25 | 26 | console.log(arguments); 27 | console.log(model_path); 28 | var select = gradioApp().querySelector("#additional_networks_metadata_editor_model > label > select"); 29 | 30 | var opt = [...select.options].filter(o => o.text == model_path)[0]; 31 | if (opt == null) { 32 | return; 33 | } 34 | 35 | addnet_switch_to_metadata_editor(); 36 | select.selectedIndex = opt.index; 37 | select.dispatchEvent(new Event("change", { bubbles: true })); 38 | 39 | return args_to_array(arguments); 40 | } 41 | -------------------------------------------------------------------------------- /models/lora/.keep: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/kohya-ss/sd-webui-additional-networks/e9f3d622b5a98650008a685ea23b27eb810da35a/models/lora/.keep -------------------------------------------------------------------------------- /preload.py: -------------------------------------------------------------------------------- 1 | import os 2 | from modules import paths 3 | 4 | 5 | def preload(parser): 6 | parser.add_argument("--addnet-max-model-count", type=int, help="The maximum number of additional network model can be used.", default=5) -------------------------------------------------------------------------------- /scripts/additional_networks.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import torch 4 | import numpy as np 5 | 6 | import modules.scripts as scripts 7 | from modules import shared, script_callbacks 8 | import gradio as gr 9 | 10 | import modules.ui 11 | from modules.ui_components import ToolButton, FormRow 12 | 13 | from scripts import addnet_xyz_grid_support, lora_compvis, model_util, metadata_editor 14 | from scripts.model_util import lora_models, MAX_MODEL_COUNT 15 | 16 | 17 | memo_symbol = "\U0001F4DD" # 📝 18 | addnet_paste_params = {"txt2img": [], "img2img": []} 19 | 20 | 21 | class Script(scripts.Script): 22 | def __init__(self) -> None: 23 | super().__init__() 24 | self.latest_params = [(None, None, None, None)] * MAX_MODEL_COUNT 25 | self.latest_networks = [] 26 | self.latest_model_hash = "" 27 | 28 | def title(self): 29 | return "Additional networks for generating" 30 | 31 | def show(self, is_img2img): 32 | return scripts.AlwaysVisible 33 | 34 | def ui(self, is_img2img): 35 | global addnet_paste_params 36 | # NOTE: Changing the contents of `ctrls` means the XY Grid support may need 37 | # to be updated, see xyz_grid_support.py 38 | ctrls = [] 39 | weight_sliders = [] 40 | model_dropdowns = [] 41 | 42 | tabname = "txt2img" 43 | if is_img2img: 44 | tabname = "img2img" 45 | 46 | paste_params = addnet_paste_params[tabname] 47 | paste_params.clear() 48 | 49 | self.infotext_fields = [] 50 | self.paste_field_names = [] 51 | 52 | with gr.Group(): 53 | with gr.Accordion("Additional Networks", open=False): 54 | with gr.Row(): 55 | enabled = gr.Checkbox(label="Enable", value=False) 56 | ctrls.append(enabled) 57 | self.infotext_fields.append((enabled, "AddNet Enabled")) 58 | separate_weights = gr.Checkbox(label="Separate UNet/Text Encoder weights", value=False) 59 | ctrls.append(separate_weights) 60 | self.infotext_fields.append((separate_weights, "AddNet Separate Weights")) 61 | 62 | for i in range(MAX_MODEL_COUNT): 63 | with FormRow(variant="compact"): 64 | module = gr.Dropdown(["LoRA"], label=f"Network module {i+1}", value="LoRA") 65 | model = gr.Dropdown(list(lora_models.keys()), label=f"Model {i+1}", value="None") 66 | with gr.Row(visible=False): 67 | model_path = gr.Textbox(value="None", interactive=False, visible=False) 68 | model.change( 69 | lambda module, model, i=i: model_util.lora_models.get(model, "None"), 70 | inputs=[module, model], 71 | outputs=[model_path], 72 | ) 73 | 74 | # Sending from the script UI to the metadata editor has to bypass 75 | # gradio since this button will exit the gr.Blocks context by the 76 | # time the metadata editor tab is created, so event handlers can't 77 | # be registered on it by then. 78 | model_info = ToolButton(value=memo_symbol, elem_id=f"additional_networks_send_to_metadata_editor_{i}") 79 | model_info.click(fn=None, _js="addnet_send_to_metadata_editor", inputs=[module, model_path], outputs=[]) 80 | 81 | module.change( 82 | lambda module, model, i=i: addnet_xyz_grid_support.update_axis_params(i, module, model), 83 | inputs=[module, model], 84 | outputs=[], 85 | ) 86 | model.change( 87 | lambda module, model, i=i: addnet_xyz_grid_support.update_axis_params(i, module, model), 88 | inputs=[module, model], 89 | outputs=[], 90 | ) 91 | 92 | # perhaps there is no user to train Text Encoder only, Weight A is U-Net 93 | # The name of label will be changed in future (Weight A and B), but UNet and TEnc for now for easy understanding 94 | with gr.Column() as col: 95 | weight = gr.Slider(label=f"Weight {i+1}", value=1.0, minimum=-1.0, maximum=2.0, step=0.05, visible=True) 96 | weight_unet = gr.Slider( 97 | label=f"UNet Weight {i+1}", value=1.0, minimum=-1.0, maximum=2.0, step=0.05, visible=False 98 | ) 99 | weight_tenc = gr.Slider( 100 | label=f"TEnc Weight {i+1}", value=1.0, minimum=-1.0, maximum=2.0, step=0.05, visible=False 101 | ) 102 | 103 | weight.change(lambda w: (w, w), inputs=[weight], outputs=[weight_unet, weight_tenc]) 104 | weight.release(lambda w: (w, w), inputs=[weight], outputs=[weight_unet, weight_tenc]) 105 | paste_params.append({"module": module, "model": model}) 106 | 107 | ctrls.extend((module, model, weight_unet, weight_tenc)) 108 | weight_sliders.extend((weight, weight_unet, weight_tenc)) 109 | model_dropdowns.append(model) 110 | 111 | self.infotext_fields.extend( 112 | [ 113 | (module, f"AddNet Module {i+1}"), 114 | (model, f"AddNet Model {i+1}"), 115 | (weight, f"AddNet Weight {i+1}"), 116 | (weight_unet, f"AddNet Weight A {i+1}"), 117 | (weight_tenc, f"AddNet Weight B {i+1}"), 118 | ] 119 | ) 120 | 121 | for _, field_name in self.infotext_fields: 122 | self.paste_field_names.append(field_name) 123 | 124 | def update_weight_sliders(separate, *sliders): 125 | updates = [] 126 | for w, w_unet, w_tenc in zip(*(iter(sliders),) * 3): 127 | if not separate: 128 | w_unet = w 129 | w_tenc = w 130 | updates.append(gr.Slider.update(visible=not separate)) # Combined 131 | updates.append(gr.Slider.update(visible=separate, value=w_unet)) # UNet 132 | updates.append(gr.Slider.update(visible=separate, value=w_tenc)) # TEnc 133 | return updates 134 | 135 | separate_weights.change(update_weight_sliders, inputs=[separate_weights] + weight_sliders, outputs=weight_sliders) 136 | 137 | def refresh_all_models(*dropdowns): 138 | model_util.update_models() 139 | updates = [] 140 | for dd in dropdowns: 141 | if dd in lora_models: 142 | selected = dd 143 | else: 144 | selected = "None" 145 | update = gr.Dropdown.update(value=selected, choices=list(lora_models.keys())) 146 | updates.append(update) 147 | return updates 148 | 149 | # mask for regions 150 | with gr.Accordion("Extra args", open=False): 151 | with gr.Row(): 152 | mask_image = gr.Image(label="mask image:") 153 | ctrls.append(mask_image) 154 | 155 | refresh_models = gr.Button(value="Refresh models") 156 | refresh_models.click(refresh_all_models, inputs=model_dropdowns, outputs=model_dropdowns) 157 | ctrls.append(refresh_models) 158 | 159 | return ctrls 160 | 161 | def set_infotext_fields(self, p, params): 162 | for i, t in enumerate(params): 163 | module, model, weight_unet, weight_tenc = t 164 | if model is None or model == "None" or len(model) == 0 or (weight_unet == 0 and weight_tenc == 0): 165 | continue 166 | p.extra_generation_params.update( 167 | { 168 | "AddNet Enabled": True, 169 | f"AddNet Module {i+1}": module, 170 | f"AddNet Model {i+1}": model, 171 | f"AddNet Weight A {i+1}": weight_unet, 172 | f"AddNet Weight B {i+1}": weight_tenc, 173 | } 174 | ) 175 | 176 | def restore_networks(self, sd_model): 177 | unet = sd_model.model.diffusion_model 178 | text_encoder = sd_model.cond_stage_model 179 | 180 | if len(self.latest_networks) > 0: 181 | print("restoring last networks") 182 | for network, _ in self.latest_networks[::-1]: 183 | network.restore(text_encoder, unet) 184 | self.latest_networks.clear() 185 | 186 | def process_batch(self, p, *args, **kwargs): 187 | unet = p.sd_model.model.diffusion_model 188 | text_encoder = p.sd_model.cond_stage_model 189 | 190 | if not args[0]: 191 | self.restore_networks(p.sd_model) 192 | return 193 | 194 | params = [] 195 | for i, ctrl in enumerate(args[2:]): 196 | if i % 4 == 0: 197 | param = [ctrl] 198 | else: 199 | param.append(ctrl) 200 | if i % 4 == 3: 201 | params.append(param) 202 | 203 | models_changed = len(self.latest_networks) == 0 # no latest network (cleared by check-off) 204 | models_changed = models_changed or self.latest_model_hash != p.sd_model.sd_model_hash 205 | if not models_changed: 206 | for (l_module, l_model, l_weight_unet, l_weight_tenc), (module, model, weight_unet, weight_tenc) in zip( 207 | self.latest_params, params 208 | ): 209 | if l_module != module or l_model != model or l_weight_unet != weight_unet or l_weight_tenc != weight_tenc: 210 | models_changed = True 211 | break 212 | 213 | if models_changed: 214 | self.restore_networks(p.sd_model) 215 | self.latest_params = params 216 | self.latest_model_hash = p.sd_model.sd_model_hash 217 | 218 | for module, model, weight_unet, weight_tenc in self.latest_params: 219 | if model is None or model == "None" or len(model) == 0: 220 | continue 221 | if weight_unet == 0 and weight_tenc == 0: 222 | print(f"ignore because weight is 0: {model}") 223 | continue 224 | 225 | model_path = lora_models.get(model, None) 226 | if model_path is None: 227 | raise RuntimeError(f"model not found: {model}") 228 | 229 | if model_path.startswith('"') and model_path.endswith('"'): # trim '"' at start/end 230 | model_path = model_path[1:-1] 231 | if not os.path.exists(model_path): 232 | print(f"file not found: {model_path}") 233 | continue 234 | 235 | print(f"{module} weight_unet: {weight_unet}, weight_tenc: {weight_tenc}, model: {model}") 236 | if module == "LoRA": 237 | if os.path.splitext(model_path)[1] == ".safetensors": 238 | from safetensors.torch import load_file 239 | 240 | du_state_dict = load_file(model_path) 241 | else: 242 | du_state_dict = torch.load(model_path, map_location="cpu") 243 | 244 | network, info = lora_compvis.create_network_and_apply_compvis( 245 | du_state_dict, weight_tenc, weight_unet, text_encoder, unet 246 | ) 247 | # in medvram, device is different for u-net and sd_model, so use sd_model's 248 | network.to(p.sd_model.device, dtype=p.sd_model.dtype) 249 | 250 | print(f"LoRA model {model} loaded: {info}") 251 | self.latest_networks.append((network, model)) 252 | if len(self.latest_networks) > 0: 253 | print("setting (or sd model) changed. new networks created.") 254 | 255 | # apply mask: currently only top 3 networks are supported 256 | if len(self.latest_networks) > 0: 257 | mask_image = args[-2] 258 | if mask_image is not None: 259 | mask_image = mask_image.astype(np.float32) / 255.0 260 | print(f"use mask image to control LoRA regions.") 261 | for i, (network, model) in enumerate(self.latest_networks[:3]): 262 | if not hasattr(network, "set_mask"): 263 | continue 264 | mask = mask_image[:, :, i] # R,G,B 265 | if mask.max() <= 0: 266 | continue 267 | mask = torch.tensor(mask, dtype=p.sd_model.dtype, device=p.sd_model.device) 268 | 269 | network.set_mask(mask, height=p.height, width=p.width, hr_height=p.hr_upscale_to_y, hr_width=p.hr_upscale_to_x) 270 | print(f"apply mask. channel: {i}, model: {model}") 271 | else: 272 | for network, _ in self.latest_networks: 273 | if hasattr(network, "set_mask"): 274 | network.set_mask(None) 275 | 276 | self.set_infotext_fields(p, self.latest_params) 277 | 278 | 279 | def on_script_unloaded(): 280 | if shared.sd_model: 281 | for s in scripts.scripts_txt2img.alwayson_scripts: 282 | if isinstance(s, Script): 283 | s.restore_networks(shared.sd_model) 284 | break 285 | 286 | 287 | def on_ui_tabs(): 288 | global addnet_paste_params 289 | with gr.Blocks(analytics_enabled=False) as additional_networks_interface: 290 | metadata_editor.setup_ui(addnet_paste_params) 291 | 292 | return [(additional_networks_interface, "Additional Networks", "additional_networks")] 293 | 294 | 295 | def on_ui_settings(): 296 | section = ("additional_networks", "Additional Networks") 297 | shared.opts.add_option( 298 | "additional_networks_extra_lora_path", 299 | shared.OptionInfo( 300 | "", 301 | """Extra paths to scan for LoRA models, comma-separated. Paths containing commas must be enclosed in double quotes. In the path, " (one quote) must be replaced by "" (two quotes).""", 302 | section=section, 303 | ), 304 | ) 305 | shared.opts.add_option( 306 | "additional_networks_sort_models_by", 307 | shared.OptionInfo( 308 | "name", 309 | "Sort LoRA models by", 310 | gr.Radio, 311 | {"choices": ["name", "date", "path name", "rating", "has user metadata"]}, 312 | section=section, 313 | ), 314 | ) 315 | shared.opts.add_option( 316 | "additional_networks_reverse_sort_order", shared.OptionInfo(False, "Reverse model sort order", section=section) 317 | ) 318 | shared.opts.add_option( 319 | "additional_networks_model_name_filter", shared.OptionInfo("", "LoRA model name filter", section=section) 320 | ) 321 | shared.opts.add_option( 322 | "additional_networks_xy_grid_model_metadata", 323 | shared.OptionInfo( 324 | "", 325 | 'Metadata to show in XY-Grid label for Model axes, comma-separated (example: "ss_learning_rate, ss_num_epochs")', 326 | section=section, 327 | ), 328 | ) 329 | shared.opts.add_option( 330 | "additional_networks_hash_thread_count", 331 | shared.OptionInfo(1, "# of threads to use for hash calculation (increase if using an SSD)", section=section), 332 | ) 333 | shared.opts.add_option( 334 | "additional_networks_back_up_model_when_saving", 335 | shared.OptionInfo(True, "Make a backup copy of the model being edited when saving its metadata.", section=section), 336 | ) 337 | shared.opts.add_option( 338 | "additional_networks_show_only_safetensors", 339 | shared.OptionInfo(False, "Only show .safetensors format models", section=section), 340 | ) 341 | shared.opts.add_option( 342 | "additional_networks_show_only_models_with_metadata", 343 | shared.OptionInfo( 344 | "disabled", 345 | "Only show models that have/don't have user-added metadata", 346 | gr.Radio, 347 | {"choices": ["disabled", "has metadata", "missing metadata"]}, 348 | section=section, 349 | ), 350 | ) 351 | shared.opts.add_option( 352 | "additional_networks_max_top_tags", shared.OptionInfo(20, "Max number of top tags to show", section=section) 353 | ) 354 | shared.opts.add_option( 355 | "additional_networks_max_dataset_folders", shared.OptionInfo(20, "Max number of dataset folders to show", section=section) 356 | ) 357 | 358 | 359 | def on_infotext_pasted(infotext, params): 360 | if "AddNet Enabled" not in params: 361 | params["AddNet Enabled"] = "False" 362 | 363 | # TODO changing "AddNet Separate Weights" does not seem to work 364 | if "AddNet Separate Weights" not in params: 365 | params["AddNet Separate Weights"] = "False" 366 | 367 | for i in range(MAX_MODEL_COUNT): 368 | # Convert combined weight into new format 369 | if f"AddNet Weight {i+1}" in params: 370 | params[f"AddNet Weight A {i+1}"] = params[f"AddNet Weight {i+1}"] 371 | params[f"AddNet Weight B {i+1}"] = params[f"AddNet Weight {i+1}"] 372 | 373 | if f"AddNet Module {i+1}" not in params: 374 | params[f"AddNet Module {i+1}"] = "LoRA" 375 | if f"AddNet Model {i+1}" not in params: 376 | params[f"AddNet Model {i+1}"] = "None" 377 | if f"AddNet Weight A {i+1}" not in params: 378 | params[f"AddNet Weight A {i+1}"] = "0" 379 | if f"AddNet Weight B {i+1}" not in params: 380 | params[f"AddNet Weight B {i+1}"] = "0" 381 | 382 | params[f"AddNet Weight {i+1}"] = params[f"AddNet Weight A {i+1}"] 383 | 384 | if params[f"AddNet Weight A {i+1}"] != params[f"AddNet Weight B {i+1}"]: 385 | params["AddNet Separate Weights"] = "True" 386 | 387 | # Convert potential legacy name/hash to new format 388 | params[f"AddNet Model {i+1}"] = str(model_util.find_closest_lora_model_name(params[f"AddNet Model {i+1}"])) 389 | 390 | addnet_xyz_grid_support.update_axis_params(i, params[f"AddNet Module {i+1}"], params[f"AddNet Model {i+1}"]) 391 | 392 | 393 | addnet_xyz_grid_support.initialize(Script) 394 | 395 | 396 | script_callbacks.on_script_unloaded(on_script_unloaded) 397 | script_callbacks.on_ui_tabs(on_ui_tabs) 398 | script_callbacks.on_ui_settings(on_ui_settings) 399 | script_callbacks.on_infotext_pasted(on_infotext_pasted) 400 | -------------------------------------------------------------------------------- /scripts/addnet_xyz_grid_support.py: -------------------------------------------------------------------------------- 1 | import os 2 | import os.path 3 | from modules import shared 4 | import modules.scripts as scripts 5 | from scripts import model_util, util 6 | from scripts.model_util import MAX_MODEL_COUNT 7 | 8 | 9 | LORA_TRAIN_METADATA_NAMES = { 10 | "ss_session_id": "Session ID", 11 | "ss_training_started_at": "Training started at", 12 | "ss_output_name": "Output name", 13 | "ss_learning_rate": "Learning rate", 14 | "ss_text_encoder_lr": "Text encoder LR", 15 | "ss_unet_lr": "UNet LR", 16 | "ss_num_train_images": "# of training images", 17 | "ss_num_reg_images": "# of reg images", 18 | "ss_num_batches_per_epoch": "Batches per epoch", 19 | "ss_num_epochs": "Total epochs", 20 | "ss_epoch": "Epoch", 21 | "ss_batch_size_per_device": "Batch size/device", 22 | "ss_total_batch_size": "Total batch size", 23 | "ss_gradient_checkpointing": "Gradient checkpointing", 24 | "ss_gradient_accumulation_steps": "Gradient accum. steps", 25 | "ss_max_train_steps": "Max train steps", 26 | "ss_lr_warmup_steps": "LR warmup steps", 27 | "ss_lr_scheduler": "LR scheduler", 28 | "ss_network_module": "Network module", 29 | "ss_network_dim": "Network dim", 30 | "ss_network_alpha": "Network alpha", 31 | "ss_mixed_precision": "Mixed precision", 32 | "ss_full_fp16": "Full FP16", 33 | "ss_v2": "V2", 34 | "ss_resolution": "Resolution", 35 | "ss_clip_skip": "Clip skip", 36 | "ss_max_token_length": "Max token length", 37 | "ss_color_aug": "Color aug", 38 | "ss_flip_aug": "Flip aug", 39 | "ss_random_crop": "Random crop", 40 | "ss_shuffle_caption": "Shuffle caption", 41 | "ss_cache_latents": "Cache latents", 42 | "ss_enable_bucket": "Enable bucket", 43 | "ss_min_bucket_reso": "Min bucket reso.", 44 | "ss_max_bucket_reso": "Max bucket reso.", 45 | "ss_seed": "Seed", 46 | "ss_keep_tokens": "Keep tokens", 47 | "ss_dataset_dirs": "Dataset dirs.", 48 | "ss_reg_dataset_dirs": "Reg dataset dirs.", 49 | "ss_sd_model_name": "SD model name", 50 | "ss_vae_name": "VAE name", 51 | "ss_training_comment": "Comment", 52 | } 53 | 54 | 55 | xy_grid = None # XY Grid module 56 | script_class = None # additional_networks scripts.Script class 57 | axis_params = [{}] * MAX_MODEL_COUNT 58 | 59 | 60 | def update_axis_params(i, module, model): 61 | axis_params[i] = {"module": module, "model": model} 62 | 63 | 64 | def get_axis_model_choices(i): 65 | module = axis_params[i].get("module", "None") 66 | model = axis_params[i].get("model", "None") 67 | 68 | if module == "LoRA": 69 | if model != "None": 70 | sort_by = shared.opts.data.get("additional_networks_sort_models_by", "name") 71 | return ["None"] + model_util.get_model_list(module, model, "", sort_by) 72 | 73 | return [f"select `Model {i+1}` in `Additional Networks`. models in same folder for selected one will be shown here."] 74 | 75 | 76 | def update_script_args(p, value, arg_idx): 77 | global script_class 78 | for s in scripts.scripts_txt2img.alwayson_scripts: 79 | if isinstance(s, script_class): 80 | args = list(p.script_args) 81 | # print(f"Changed arg {arg_idx} from {args[s.args_from + arg_idx - 1]} to {value}") 82 | args[s.args_from + arg_idx] = value 83 | p.script_args = tuple(args) 84 | break 85 | 86 | 87 | def confirm_models(p, xs): 88 | for x in xs: 89 | if x in ["", "None"]: 90 | continue 91 | if not model_util.find_closest_lora_model_name(x): 92 | raise RuntimeError(f"Unknown LoRA model: {x}") 93 | 94 | 95 | def apply_module(p, x, xs, i): 96 | update_script_args(p, True, 0) # set Enabled to True 97 | update_script_args(p, x, 2 + 4 * i) # enabled, separate_weights, ({module}, model, weight_unet, weight_tenc), ... 98 | 99 | 100 | def apply_model(p, x, xs, i): 101 | name = model_util.find_closest_lora_model_name(x) 102 | update_script_args(p, True, 0) 103 | update_script_args(p, name, 3 + 4 * i) # enabled, separate_weights, (module, {model}, weight_unet, weight_tenc), ... 104 | 105 | 106 | def apply_weight(p, x, xs, i): 107 | update_script_args(p, True, 0) 108 | update_script_args(p, x, 4 + 4 * i) # enabled, separate_weights, (module, model, {weight_unet, weight_tenc}), ... 109 | update_script_args(p, x, 5 + 4 * i) 110 | 111 | 112 | def apply_weight_unet(p, x, xs, i): 113 | update_script_args(p, True, 0) 114 | update_script_args(p, x, 4 + 4 * i) # enabled, separate_weights, (module, model, {weight_unet}, weight_tenc), ... 115 | 116 | 117 | def apply_weight_tenc(p, x, xs, i): 118 | update_script_args(p, True, 0) 119 | update_script_args(p, x, 5 + 4 * i) # enabled, separate_weights, (module, model, weight_unet, {weight_tenc}), ... 120 | 121 | 122 | def format_lora_model(p, opt, x): 123 | global xy_grid 124 | model = model_util.find_closest_lora_model_name(x) 125 | if model is None or model.lower() in ["", "none"]: 126 | return "None" 127 | 128 | value = xy_grid.format_value(p, opt, model) 129 | 130 | model_path = model_util.lora_models.get(model) 131 | metadata = model_util.read_model_metadata(model_path, "LoRA") 132 | if not metadata: 133 | return value 134 | 135 | metadata_names = util.split_path_list(shared.opts.data.get("additional_networks_xy_grid_model_metadata", "")) 136 | if not metadata_names: 137 | return value 138 | 139 | for name in metadata_names: 140 | name = name.strip() 141 | if name in metadata: 142 | formatted_name = LORA_TRAIN_METADATA_NAMES.get(name, name) 143 | value += f"\n{formatted_name}: {metadata[name]}, " 144 | 145 | return value.strip(" ").strip(",") 146 | 147 | 148 | def initialize(script): 149 | global xy_grid, script_class 150 | xy_grid = None 151 | script_class = script 152 | for scriptDataTuple in scripts.scripts_data: 153 | if os.path.basename(scriptDataTuple.path) == "xy_grid.py" or os.path.basename(scriptDataTuple.path) == "xyz_grid.py": 154 | xy_grid = scriptDataTuple.module 155 | for i in range(MAX_MODEL_COUNT): 156 | model = xy_grid.AxisOption( 157 | f"AddNet Model {i+1}", 158 | str, 159 | lambda p, x, xs, i=i: apply_model(p, x, xs, i), 160 | format_lora_model, 161 | confirm_models, 162 | cost=0.5, 163 | choices=lambda i=i: get_axis_model_choices(i), 164 | ) 165 | weight = xy_grid.AxisOption( 166 | f"AddNet Weight {i+1}", 167 | float, 168 | lambda p, x, xs, i=i: apply_weight(p, x, xs, i), 169 | xy_grid.format_value_add_label, 170 | None, 171 | cost=0.5, 172 | ) 173 | weight_unet = xy_grid.AxisOption( 174 | f"AddNet UNet Weight {i+1}", 175 | float, 176 | lambda p, x, xs, i=i: apply_weight_unet(p, x, xs, i), 177 | xy_grid.format_value_add_label, 178 | None, 179 | cost=0.5, 180 | ) 181 | weight_tenc = xy_grid.AxisOption( 182 | f"AddNet TEnc Weight {i+1}", 183 | float, 184 | lambda p, x, xs, i=i: apply_weight_tenc(p, x, xs, i), 185 | xy_grid.format_value_add_label, 186 | None, 187 | cost=0.5, 188 | ) 189 | xy_grid.axis_options.extend([model, weight, weight_unet, weight_tenc]) 190 | -------------------------------------------------------------------------------- /scripts/lora_compvis.py: -------------------------------------------------------------------------------- 1 | # LoRA network module 2 | # reference: 3 | # https://github.com/microsoft/LoRA/blob/main/loralib/layers.py 4 | # https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py 5 | 6 | import copy 7 | import math 8 | import re 9 | from typing import NamedTuple 10 | import torch 11 | 12 | 13 | class LoRAInfo(NamedTuple): 14 | lora_name: str 15 | module_name: str 16 | module: torch.nn.Module 17 | multiplier: float 18 | dim: int 19 | alpha: float 20 | 21 | 22 | class LoRAModule(torch.nn.Module): 23 | """ 24 | replaces forward method of the original Linear, instead of replacing the original Linear module. 25 | """ 26 | 27 | def __init__(self, lora_name, org_module: torch.nn.Module, multiplier=1.0, lora_dim=4, alpha=1): 28 | """if alpha == 0 or None, alpha is rank (no scaling).""" 29 | super().__init__() 30 | self.lora_name = lora_name 31 | self.lora_dim = lora_dim 32 | 33 | if org_module.__class__.__name__ == "Conv2d": 34 | in_dim = org_module.in_channels 35 | out_dim = org_module.out_channels 36 | 37 | # self.lora_dim = min(self.lora_dim, in_dim, out_dim) 38 | # if self.lora_dim != lora_dim: 39 | # print(f"{lora_name} dim (rank) is changed to: {self.lora_dim}") 40 | 41 | kernel_size = org_module.kernel_size 42 | stride = org_module.stride 43 | padding = org_module.padding 44 | self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False) 45 | self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False) 46 | else: 47 | in_dim = org_module.in_features 48 | out_dim = org_module.out_features 49 | self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False) 50 | self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False) 51 | 52 | if type(alpha) == torch.Tensor: 53 | alpha = alpha.detach().float().numpy() # without casting, bf16 causes error 54 | alpha = self.lora_dim if alpha is None or alpha == 0 else alpha 55 | self.scale = alpha / self.lora_dim 56 | self.register_buffer("alpha", torch.tensor(alpha)) # 定数として扱える 57 | 58 | # same as microsoft's 59 | torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5)) 60 | torch.nn.init.zeros_(self.lora_up.weight) 61 | 62 | self.multiplier = multiplier 63 | self.org_forward = org_module.forward 64 | self.org_module = org_module # remove in applying 65 | self.mask_dic = None 66 | self.mask = None 67 | self.mask_area = -1 68 | 69 | def apply_to(self): 70 | self.org_forward = self.org_module.forward 71 | self.org_module.forward = self.forward 72 | del self.org_module 73 | 74 | def set_mask_dic(self, mask_dic): 75 | # called before every generation 76 | 77 | # check this module is related to h,w (not context and time emb) 78 | if "attn2_to_k" in self.lora_name or "attn2_to_v" in self.lora_name or "emb_layers" in self.lora_name: 79 | # print(f"LoRA for context or time emb: {self.lora_name}") 80 | self.mask_dic = None 81 | else: 82 | self.mask_dic = mask_dic 83 | 84 | self.mask = None 85 | 86 | def forward(self, x): 87 | """ 88 | may be cascaded. 89 | """ 90 | if self.mask_dic is None: 91 | return self.org_forward(x) + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale 92 | 93 | # regional LoRA 94 | 95 | # calculate lora and get size 96 | lx = self.lora_up(self.lora_down(x)) 97 | 98 | if len(lx.size()) == 4: # b,c,h,w 99 | area = lx.size()[2] * lx.size()[3] 100 | else: 101 | area = lx.size()[1] # b,seq,dim 102 | 103 | if self.mask is None or self.mask_area != area: 104 | # get mask 105 | # print(self.lora_name, x.size(), lx.size(), area) 106 | mask = self.mask_dic[area] 107 | if len(lx.size()) == 3: 108 | mask = torch.reshape(mask, (1, -1, 1)) 109 | self.mask = mask 110 | self.mask_area = area 111 | 112 | return self.org_forward(x) + lx * self.multiplier * self.scale * self.mask 113 | 114 | 115 | def create_network_and_apply_compvis(du_state_dict, multiplier_tenc, multiplier_unet, text_encoder, unet, **kwargs): 116 | # get device and dtype from unet 117 | for module in unet.modules(): 118 | if module.__class__.__name__ == "Linear": 119 | param: torch.nn.Parameter = module.weight 120 | # device = param.device 121 | dtype = param.dtype 122 | break 123 | 124 | # get dims (rank) and alpha from state dict 125 | modules_dim = {} 126 | modules_alpha = {} 127 | for key, value in du_state_dict.items(): 128 | if "." not in key: 129 | continue 130 | 131 | lora_name = key.split(".")[0] 132 | if "alpha" in key: 133 | modules_alpha[lora_name] = float(value.detach().to(torch.float).cpu().numpy()) 134 | elif "lora_down" in key: 135 | dim = value.size()[0] 136 | modules_dim[lora_name] = dim 137 | 138 | # support old LoRA without alpha 139 | for key in modules_dim.keys(): 140 | if key not in modules_alpha: 141 | modules_alpha[key] = modules_dim[key] 142 | 143 | print( 144 | f"dimension: {set(modules_dim.values())}, alpha: {set(modules_alpha.values())}, multiplier_unet: {multiplier_unet}, multiplier_tenc: {multiplier_tenc}" 145 | ) 146 | 147 | # if network_dim is None: 148 | # print(f"The selected model is not LoRA or not trained by `sd-scripts`?") 149 | # network_dim = 4 150 | # network_alpha = 1 151 | 152 | # create, apply and load weights 153 | network = LoRANetworkCompvis(text_encoder, unet, multiplier_tenc, multiplier_unet, modules_dim, modules_alpha) 154 | state_dict = network.apply_lora_modules(du_state_dict) # some weights are applied to text encoder 155 | network.to(dtype) # with this, if error comes from next line, the model will be used 156 | info = network.load_state_dict(state_dict, strict=False) 157 | 158 | # remove redundant warnings 159 | if len(info.missing_keys) > 4: 160 | missing_keys = [] 161 | alpha_count = 0 162 | for key in info.missing_keys: 163 | if "alpha" not in key: 164 | missing_keys.append(key) 165 | else: 166 | if alpha_count == 0: 167 | missing_keys.append(key) 168 | alpha_count += 1 169 | if alpha_count > 1: 170 | missing_keys.append( 171 | f"... and {alpha_count-1} alphas. The model doesn't have alpha, use dim (rannk) as alpha. You can ignore this message." 172 | ) 173 | 174 | info = torch.nn.modules.module._IncompatibleKeys(missing_keys, info.unexpected_keys) 175 | 176 | return network, info 177 | 178 | 179 | class LoRANetworkCompvis(torch.nn.Module): 180 | # UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel", "Attention"] 181 | # TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPMLP"] 182 | UNET_TARGET_REPLACE_MODULE = ["SpatialTransformer", "ResBlock", "Downsample", "Upsample"] # , "Attention"] 183 | TEXT_ENCODER_TARGET_REPLACE_MODULE = ["ResidualAttentionBlock", "CLIPAttention", "CLIPMLP"] 184 | 185 | LORA_PREFIX_UNET = "lora_unet" 186 | LORA_PREFIX_TEXT_ENCODER = "lora_te" 187 | 188 | @classmethod 189 | def convert_diffusers_name_to_compvis(cls, v2, du_name): 190 | """ 191 | convert diffusers's LoRA name to CompVis 192 | """ 193 | cv_name = None 194 | if "lora_unet_" in du_name: 195 | m = re.search(r"_down_blocks_(\d+)_attentions_(\d+)_(.+)", du_name) 196 | if m: 197 | du_block_index = int(m.group(1)) 198 | du_attn_index = int(m.group(2)) 199 | du_suffix = m.group(3) 200 | 201 | cv_index = 1 + du_block_index * 3 + du_attn_index # 1,2, 4,5, 7,8 202 | cv_name = f"lora_unet_input_blocks_{cv_index}_1_{du_suffix}" 203 | return cv_name 204 | 205 | m = re.search(r"_mid_block_attentions_(\d+)_(.+)", du_name) 206 | if m: 207 | du_suffix = m.group(2) 208 | cv_name = f"lora_unet_middle_block_1_{du_suffix}" 209 | return cv_name 210 | 211 | m = re.search(r"_up_blocks_(\d+)_attentions_(\d+)_(.+)", du_name) 212 | if m: 213 | du_block_index = int(m.group(1)) 214 | du_attn_index = int(m.group(2)) 215 | du_suffix = m.group(3) 216 | 217 | cv_index = du_block_index * 3 + du_attn_index # 3,4,5, 6,7,8, 9,10,11 218 | cv_name = f"lora_unet_output_blocks_{cv_index}_1_{du_suffix}" 219 | return cv_name 220 | 221 | m = re.search(r"_down_blocks_(\d+)_resnets_(\d+)_(.+)", du_name) 222 | if m: 223 | du_block_index = int(m.group(1)) 224 | du_res_index = int(m.group(2)) 225 | du_suffix = m.group(3) 226 | cv_suffix = { 227 | "conv1": "in_layers_2", 228 | "conv2": "out_layers_3", 229 | "time_emb_proj": "emb_layers_1", 230 | "conv_shortcut": "skip_connection", 231 | }[du_suffix] 232 | 233 | cv_index = 1 + du_block_index * 3 + du_res_index # 1,2, 4,5, 7,8 234 | cv_name = f"lora_unet_input_blocks_{cv_index}_0_{cv_suffix}" 235 | return cv_name 236 | 237 | m = re.search(r"_down_blocks_(\d+)_downsamplers_0_conv", du_name) 238 | if m: 239 | block_index = int(m.group(1)) 240 | cv_index = 3 + block_index * 3 241 | cv_name = f"lora_unet_input_blocks_{cv_index}_0_op" 242 | return cv_name 243 | 244 | m = re.search(r"_mid_block_resnets_(\d+)_(.+)", du_name) 245 | if m: 246 | index = int(m.group(1)) 247 | du_suffix = m.group(2) 248 | cv_suffix = { 249 | "conv1": "in_layers_2", 250 | "conv2": "out_layers_3", 251 | "time_emb_proj": "emb_layers_1", 252 | "conv_shortcut": "skip_connection", 253 | }[du_suffix] 254 | cv_name = f"lora_unet_middle_block_{index*2}_{cv_suffix}" 255 | return cv_name 256 | 257 | m = re.search(r"_up_blocks_(\d+)_resnets_(\d+)_(.+)", du_name) 258 | if m: 259 | du_block_index = int(m.group(1)) 260 | du_res_index = int(m.group(2)) 261 | du_suffix = m.group(3) 262 | cv_suffix = { 263 | "conv1": "in_layers_2", 264 | "conv2": "out_layers_3", 265 | "time_emb_proj": "emb_layers_1", 266 | "conv_shortcut": "skip_connection", 267 | }[du_suffix] 268 | 269 | cv_index = du_block_index * 3 + du_res_index # 1,2, 4,5, 7,8 270 | cv_name = f"lora_unet_output_blocks_{cv_index}_0_{cv_suffix}" 271 | return cv_name 272 | 273 | m = re.search(r"_up_blocks_(\d+)_upsamplers_0_conv", du_name) 274 | if m: 275 | block_index = int(m.group(1)) 276 | cv_index = block_index * 3 + 2 277 | cv_name = f"lora_unet_output_blocks_{cv_index}_{bool(block_index)+1}_conv" 278 | return cv_name 279 | 280 | elif "lora_te_" in du_name: 281 | m = re.search(r"_model_encoder_layers_(\d+)_(.+)", du_name) 282 | if m: 283 | du_block_index = int(m.group(1)) 284 | du_suffix = m.group(2) 285 | 286 | cv_index = du_block_index 287 | if v2: 288 | if "mlp_fc1" in du_suffix: 289 | cv_name = ( 290 | f"lora_te_wrapped_model_transformer_resblocks_{cv_index}_{du_suffix.replace('mlp_fc1', 'mlp_c_fc')}" 291 | ) 292 | elif "mlp_fc2" in du_suffix: 293 | cv_name = ( 294 | f"lora_te_wrapped_model_transformer_resblocks_{cv_index}_{du_suffix.replace('mlp_fc2', 'mlp_c_proj')}" 295 | ) 296 | elif "self_attn": 297 | # handled later 298 | cv_name = f"lora_te_wrapped_model_transformer_resblocks_{cv_index}_{du_suffix.replace('self_attn', 'attn')}" 299 | else: 300 | cv_name = f"lora_te_wrapped_transformer_text_model_encoder_layers_{cv_index}_{du_suffix}" 301 | 302 | assert cv_name is not None, f"conversion failed: {du_name}. the model may not be trained by `sd-scripts`." 303 | return cv_name 304 | 305 | @classmethod 306 | def convert_state_dict_name_to_compvis(cls, v2, state_dict): 307 | """ 308 | convert keys in state dict to load it by load_state_dict 309 | """ 310 | new_sd = {} 311 | for key, value in state_dict.items(): 312 | tokens = key.split(".") 313 | compvis_name = LoRANetworkCompvis.convert_diffusers_name_to_compvis(v2, tokens[0]) 314 | new_key = compvis_name + "." + ".".join(tokens[1:]) 315 | 316 | new_sd[new_key] = value 317 | 318 | return new_sd 319 | 320 | def __init__(self, text_encoder, unet, multiplier_tenc=1.0, multiplier_unet=1.0, modules_dim=None, modules_alpha=None) -> None: 321 | super().__init__() 322 | self.multiplier_unet = multiplier_unet 323 | self.multiplier_tenc = multiplier_tenc 324 | self.latest_mask_info = None 325 | 326 | # check v1 or v2 327 | self.v2 = False 328 | for _, module in text_encoder.named_modules(): 329 | for _, child_module in module.named_modules(): 330 | if child_module.__class__.__name__ == "MultiheadAttention": 331 | self.v2 = True 332 | break 333 | if self.v2: 334 | break 335 | 336 | # convert lora name to CompVis and get dim and alpha 337 | comp_vis_loras_dim_alpha = {} 338 | for du_lora_name in modules_dim.keys(): 339 | dim = modules_dim[du_lora_name] 340 | alpha = modules_alpha[du_lora_name] 341 | comp_vis_lora_name = LoRANetworkCompvis.convert_diffusers_name_to_compvis(self.v2, du_lora_name) 342 | comp_vis_loras_dim_alpha[comp_vis_lora_name] = (dim, alpha) 343 | 344 | # create module instances 345 | def create_modules(prefix, root_module: torch.nn.Module, target_replace_modules, multiplier): 346 | loras = [] 347 | replaced_modules = [] 348 | for name, module in root_module.named_modules(): 349 | if module.__class__.__name__ in target_replace_modules: 350 | for child_name, child_module in module.named_modules(): 351 | # enumerate all Linear and Conv2d 352 | if child_module.__class__.__name__ == "Linear" or child_module.__class__.__name__ == "Conv2d": 353 | lora_name = prefix + "." + name + "." + child_name 354 | lora_name = lora_name.replace(".", "_") 355 | if "_resblocks_23_" in lora_name: # ignore last block in StabilityAi Text Encoder 356 | break 357 | if lora_name not in comp_vis_loras_dim_alpha: 358 | continue 359 | 360 | dim, alpha = comp_vis_loras_dim_alpha[lora_name] 361 | lora = LoRAModule(lora_name, child_module, multiplier, dim, alpha) 362 | loras.append(lora) 363 | 364 | replaced_modules.append(child_module) 365 | elif child_module.__class__.__name__ == "MultiheadAttention": 366 | # make four modules: not replacing forward method but merge weights later 367 | for suffix in ["q_proj", "k_proj", "v_proj", "out_proj"]: 368 | module_name = prefix + "." + name + "." + child_name # ~.attn 369 | module_name = module_name.replace(".", "_") 370 | if "_resblocks_23_" in module_name: # ignore last block in StabilityAi Text Encoder 371 | break 372 | 373 | lora_name = module_name + "_" + suffix 374 | if lora_name not in comp_vis_loras_dim_alpha: 375 | continue 376 | dim, alpha = comp_vis_loras_dim_alpha[lora_name] 377 | lora_info = LoRAInfo(lora_name, module_name, child_module, multiplier, dim, alpha) 378 | loras.append(lora_info) 379 | 380 | replaced_modules.append(child_module) 381 | return loras, replaced_modules 382 | 383 | self.text_encoder_loras, te_rep_modules = create_modules( 384 | LoRANetworkCompvis.LORA_PREFIX_TEXT_ENCODER, 385 | text_encoder, 386 | LoRANetworkCompvis.TEXT_ENCODER_TARGET_REPLACE_MODULE, 387 | self.multiplier_tenc, 388 | ) 389 | print(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") 390 | 391 | self.unet_loras, unet_rep_modules = create_modules( 392 | LoRANetworkCompvis.LORA_PREFIX_UNET, unet, LoRANetworkCompvis.UNET_TARGET_REPLACE_MODULE, self.multiplier_unet 393 | ) 394 | print(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") 395 | 396 | # make backup of original forward/weights, if multiple modules are applied, do in 1st module only 397 | backed_up = False # messaging purpose only 398 | for rep_module in te_rep_modules + unet_rep_modules: 399 | if ( 400 | rep_module.__class__.__name__ == "MultiheadAttention" 401 | ): # multiple MHA modules are in list, prevent to backed up forward 402 | if not hasattr(rep_module, "_lora_org_weights"): 403 | # avoid updating of original weights. state_dict is reference to original weights 404 | rep_module._lora_org_weights = copy.deepcopy(rep_module.state_dict()) 405 | backed_up = True 406 | elif not hasattr(rep_module, "_lora_org_forward"): 407 | rep_module._lora_org_forward = rep_module.forward 408 | backed_up = True 409 | if backed_up: 410 | print("original forward/weights is backed up.") 411 | 412 | # assertion 413 | names = set() 414 | for lora in self.text_encoder_loras + self.unet_loras: 415 | assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}" 416 | names.add(lora.lora_name) 417 | 418 | def restore(self, text_encoder, unet): 419 | # restore forward/weights from property for all modules 420 | restored = False # messaging purpose only 421 | modules = [] 422 | modules.extend(text_encoder.modules()) 423 | modules.extend(unet.modules()) 424 | for module in modules: 425 | if hasattr(module, "_lora_org_forward"): 426 | module.forward = module._lora_org_forward 427 | del module._lora_org_forward 428 | restored = True 429 | if hasattr( 430 | module, "_lora_org_weights" 431 | ): # module doesn't have forward and weights at same time currently, but supports it for future changing 432 | module.load_state_dict(module._lora_org_weights) 433 | del module._lora_org_weights 434 | restored = True 435 | 436 | if restored: 437 | print("original forward/weights is restored.") 438 | 439 | def apply_lora_modules(self, du_state_dict): 440 | # conversion 1st step: convert names in state_dict 441 | state_dict = LoRANetworkCompvis.convert_state_dict_name_to_compvis(self.v2, du_state_dict) 442 | 443 | # check state_dict has text_encoder or unet 444 | weights_has_text_encoder = weights_has_unet = False 445 | for key in state_dict.keys(): 446 | if key.startswith(LoRANetworkCompvis.LORA_PREFIX_TEXT_ENCODER): 447 | weights_has_text_encoder = True 448 | elif key.startswith(LoRANetworkCompvis.LORA_PREFIX_UNET): 449 | weights_has_unet = True 450 | if weights_has_text_encoder and weights_has_unet: 451 | break 452 | 453 | apply_text_encoder = weights_has_text_encoder 454 | apply_unet = weights_has_unet 455 | 456 | if apply_text_encoder: 457 | print("enable LoRA for text encoder") 458 | else: 459 | self.text_encoder_loras = [] 460 | 461 | if apply_unet: 462 | print("enable LoRA for U-Net") 463 | else: 464 | self.unet_loras = [] 465 | 466 | # add modules to network: this makes state_dict can be got from LoRANetwork 467 | mha_loras = {} 468 | for lora in self.text_encoder_loras + self.unet_loras: 469 | if type(lora) == LoRAModule: 470 | lora.apply_to() # ensure remove reference to original Linear: reference makes key of state_dict 471 | self.add_module(lora.lora_name, lora) 472 | else: 473 | # SD2.x MultiheadAttention merge weights to MHA weights 474 | lora_info: LoRAInfo = lora 475 | if lora_info.module_name not in mha_loras: 476 | mha_loras[lora_info.module_name] = {} 477 | 478 | lora_dic = mha_loras[lora_info.module_name] 479 | lora_dic[lora_info.lora_name] = lora_info 480 | if len(lora_dic) == 4: 481 | # calculate and apply 482 | module = lora_info.module 483 | module_name = lora_info.module_name 484 | w_q_dw = state_dict.get(module_name + "_q_proj.lora_down.weight") 485 | if w_q_dw is not None: # corresponding LoRA module exists 486 | w_q_up = state_dict[module_name + "_q_proj.lora_up.weight"] 487 | w_k_dw = state_dict[module_name + "_k_proj.lora_down.weight"] 488 | w_k_up = state_dict[module_name + "_k_proj.lora_up.weight"] 489 | w_v_dw = state_dict[module_name + "_v_proj.lora_down.weight"] 490 | w_v_up = state_dict[module_name + "_v_proj.lora_up.weight"] 491 | w_out_dw = state_dict[module_name + "_out_proj.lora_down.weight"] 492 | w_out_up = state_dict[module_name + "_out_proj.lora_up.weight"] 493 | q_lora_info = lora_dic[module_name + "_q_proj"] 494 | k_lora_info = lora_dic[module_name + "_k_proj"] 495 | v_lora_info = lora_dic[module_name + "_v_proj"] 496 | out_lora_info = lora_dic[module_name + "_out_proj"] 497 | 498 | sd = module.state_dict() 499 | qkv_weight = sd["in_proj_weight"] 500 | out_weight = sd["out_proj.weight"] 501 | dev = qkv_weight.device 502 | 503 | def merge_weights(l_info, weight, up_weight, down_weight): 504 | # calculate in float 505 | scale = l_info.alpha / l_info.dim 506 | dtype = weight.dtype 507 | weight = ( 508 | weight.float() 509 | + l_info.multiplier 510 | * (up_weight.to(dev, dtype=torch.float) @ down_weight.to(dev, dtype=torch.float)) 511 | * scale 512 | ) 513 | weight = weight.to(dtype) 514 | return weight 515 | 516 | q_weight, k_weight, v_weight = torch.chunk(qkv_weight, 3) 517 | if q_weight.size()[1] == w_q_up.size()[0]: 518 | q_weight = merge_weights(q_lora_info, q_weight, w_q_up, w_q_dw) 519 | k_weight = merge_weights(k_lora_info, k_weight, w_k_up, w_k_dw) 520 | v_weight = merge_weights(v_lora_info, v_weight, w_v_up, w_v_dw) 521 | qkv_weight = torch.cat([q_weight, k_weight, v_weight]) 522 | 523 | out_weight = merge_weights(out_lora_info, out_weight, w_out_up, w_out_dw) 524 | 525 | sd["in_proj_weight"] = qkv_weight.to(dev) 526 | sd["out_proj.weight"] = out_weight.to(dev) 527 | 528 | lora_info.module.load_state_dict(sd) 529 | else: 530 | # different dim, version mismatch 531 | print(f"shape of weight is different: {module_name}. SD version may be different") 532 | 533 | for t in ["q", "k", "v", "out"]: 534 | del state_dict[f"{module_name}_{t}_proj.lora_down.weight"] 535 | del state_dict[f"{module_name}_{t}_proj.lora_up.weight"] 536 | alpha_key = f"{module_name}_{t}_proj.alpha" 537 | if alpha_key in state_dict: 538 | del state_dict[alpha_key] 539 | else: 540 | # corresponding weight not exists: version mismatch 541 | pass 542 | 543 | # conversion 2nd step: convert weight's shape (and handle wrapped) 544 | state_dict = self.convert_state_dict_shape_to_compvis(state_dict) 545 | 546 | return state_dict 547 | 548 | def convert_state_dict_shape_to_compvis(self, state_dict): 549 | # shape conversion 550 | current_sd = self.state_dict() # to get target shape 551 | wrapped = False 552 | count = 0 553 | for key in list(state_dict.keys()): 554 | if key not in current_sd: 555 | continue # might be error or another version 556 | if "wrapped" in key: 557 | wrapped = True 558 | 559 | value: torch.Tensor = state_dict[key] 560 | if value.size() != current_sd[key].size(): 561 | # print(f"convert weights shape: {key}, from: {value.size()}, {len(value.size())}") 562 | count += 1 563 | if len(value.size()) == 4: 564 | value = value.squeeze(3).squeeze(2) 565 | else: 566 | value = value.unsqueeze(2).unsqueeze(3) 567 | state_dict[key] = value 568 | if tuple(value.size()) != tuple(current_sd[key].size()): 569 | print( 570 | f"weight's shape is different: {key} expected {current_sd[key].size()} found {value.size()}. SD version may be different" 571 | ) 572 | del state_dict[key] 573 | print(f"shapes for {count} weights are converted.") 574 | 575 | # convert wrapped 576 | if not wrapped: 577 | print("remove 'wrapped' from keys") 578 | for key in list(state_dict.keys()): 579 | if "_wrapped_" in key: 580 | new_key = key.replace("_wrapped_", "_") 581 | state_dict[new_key] = state_dict[key] 582 | del state_dict[key] 583 | 584 | return state_dict 585 | 586 | def set_mask(self, mask, height=None, width=None, hr_height=None, hr_width=None): 587 | if mask is None: 588 | # clear latest mask 589 | # print("clear mask") 590 | self.latest_mask_info = None 591 | for lora in self.unet_loras: 592 | lora.set_mask_dic(None) 593 | return 594 | 595 | # check mask image and h/w are same 596 | if ( 597 | self.latest_mask_info is not None 598 | and torch.equal(mask, self.latest_mask_info[0]) 599 | and (height, width, hr_height, hr_width) == self.latest_mask_info[1:] 600 | ): 601 | # print("mask not changed") 602 | return 603 | 604 | self.latest_mask_info = (mask, height, width, hr_height, hr_width) 605 | 606 | org_dtype = mask.dtype 607 | if mask.dtype == torch.bfloat16: 608 | mask = mask.to(torch.float) 609 | 610 | mask_dic = {} 611 | mask = mask.unsqueeze(0).unsqueeze(1) # b(1),c(1),h,w 612 | 613 | def resize_add(mh, mw): 614 | # print(mh, mw, mh * mw) 615 | m = torch.nn.functional.interpolate(mask, (mh, mw), mode="bilinear") # doesn't work in bf16 616 | m = m.to(org_dtype) 617 | mask_dic[mh * mw] = m 618 | 619 | for h, w in [(height, width), (hr_height, hr_width)]: 620 | if not h or not w: 621 | continue 622 | 623 | h = h // 8 624 | w = w // 8 625 | for i in range(4): 626 | resize_add(h, w) 627 | if h % 2 == 1 or w % 2 == 1: # add extra shape if h/w is not divisible by 2 628 | resize_add(h + h % 2, w + w % 2) 629 | h = (h + 1) // 2 630 | w = (w + 1) // 2 631 | 632 | for lora in self.unet_loras: 633 | lora.set_mask_dic(mask_dic) 634 | return 635 | -------------------------------------------------------------------------------- /scripts/metadata_editor.py: -------------------------------------------------------------------------------- 1 | import os 2 | import json 3 | import sys 4 | import io 5 | import base64 6 | import platform 7 | import subprocess as sp 8 | from PIL import PngImagePlugin, Image 9 | 10 | from modules import shared 11 | import gradio as gr 12 | 13 | import modules.ui 14 | from modules.ui_components import ToolButton 15 | import modules.extras 16 | import modules.generation_parameters_copypaste as parameters_copypaste 17 | 18 | from scripts import safetensors_hack, model_util 19 | from scripts.model_util import MAX_MODEL_COUNT 20 | 21 | 22 | folder_symbol = "\U0001f4c2" # 📂 23 | keycap_symbols = [ 24 | "\u0031\ufe0f\u20e3", # 1️⃣ 25 | "\u0032\ufe0f\u20e3", # 2️⃣ 26 | "\u0033\ufe0f\u20e3", # 3️⃣ 27 | "\u0034\ufe0f\u20e3", # 4️⃣ 28 | "\u0035\ufe0f\u20e3", # 5️⃣ 29 | "\u0036\ufe0f\u20e3", # 6️⃣ 30 | "\u0037\ufe0f\u20e3", # 7️⃣ 31 | "\u0038\ufe0f\u20e3", # 8️ 32 | "\u0039\ufe0f\u20e3", # 9️ 33 | "\u1f51f", # 🔟 34 | ] 35 | 36 | 37 | def write_webui_model_preview_image(model_path, image): 38 | basename, ext = os.path.splitext(model_path) 39 | preview_path = f"{basename}.png" 40 | 41 | # Copy any text-only metadata 42 | use_metadata = False 43 | metadata = PngImagePlugin.PngInfo() 44 | for key, value in image.info.items(): 45 | if isinstance(key, str) and isinstance(value, str): 46 | metadata.add_text(key, value) 47 | use_metadata = True 48 | 49 | image.save(preview_path, "PNG", pnginfo=(metadata if use_metadata else None)) 50 | 51 | 52 | def delete_webui_model_preview_image(model_path): 53 | basename, ext = os.path.splitext(model_path) 54 | preview_paths = [f"{basename}.preview.png", f"{basename}.png"] 55 | 56 | for preview_path in preview_paths: 57 | if os.path.isfile(preview_path): 58 | os.unlink(preview_path) 59 | 60 | 61 | def decode_base64_to_pil(encoding): 62 | if encoding.startswith("data:image/"): 63 | encoding = encoding.split(";")[1].split(",")[1] 64 | return Image.open(io.BytesIO(base64.b64decode(encoding))) 65 | 66 | 67 | def encode_pil_to_base64(image): 68 | with io.BytesIO() as output_bytes: 69 | # Copy any text-only metadata 70 | use_metadata = False 71 | metadata = PngImagePlugin.PngInfo() 72 | for key, value in image.info.items(): 73 | if isinstance(key, str) and isinstance(value, str): 74 | metadata.add_text(key, value) 75 | use_metadata = True 76 | 77 | image.save(output_bytes, "PNG", pnginfo=(metadata if use_metadata else None)) 78 | bytes_data = output_bytes.getvalue() 79 | return base64.b64encode(bytes_data) 80 | 81 | 82 | def open_folder(f): 83 | if not os.path.exists(f): 84 | print(f'Folder "{f}" does not exist. After you create an image, the folder will be created.') 85 | return 86 | elif not os.path.isdir(f): 87 | print( 88 | f""" 89 | WARNING 90 | An open_folder request was made with an argument that is not a folder. 91 | This could be an error or a malicious attempt to run code on your computer. 92 | Requested path was: {f} 93 | """, 94 | file=sys.stderr, 95 | ) 96 | return 97 | 98 | if not shared.cmd_opts.hide_ui_dir_config: 99 | path = os.path.normpath(f) 100 | if platform.system() == "Windows": 101 | os.startfile(path) 102 | elif platform.system() == "Darwin": 103 | sp.Popen(["open", path]) 104 | elif "microsoft-standard-WSL2" in platform.uname().release: 105 | sp.Popen(["wsl-open", path]) 106 | else: 107 | sp.Popen(["xdg-open", path]) 108 | 109 | 110 | def copy_metadata_to_all(module, model_path, copy_dir, same_session_only, missing_meta_only, cover_image): 111 | """ 112 | Given a model with metadata, copies that metadata to all models in copy_dir. 113 | 114 | :str module: Module name ("LoRA") 115 | :str model: Model key in lora_models ("MyModel(123456abcdef)") 116 | :str copy_dir: Directory to copy to 117 | :bool same_session_only: Only copy to modules with the same ss_session_id 118 | :bool missing_meta_only: Only copy to modules that are missing user metadata 119 | :Optional[Image] cover_image: Cover image to embed in the file as base64 120 | :returns: gr.HTML.update() 121 | """ 122 | if model_path == "None": 123 | return "No model selected." 124 | 125 | if not os.path.isfile(model_path): 126 | return f"Model path not found: {model_path}" 127 | 128 | model_path = os.path.realpath(model_path) 129 | 130 | if os.path.splitext(model_path)[1] != ".safetensors": 131 | return "Model is not in .safetensors format." 132 | 133 | if not os.path.isdir(copy_dir): 134 | return "Please provide a directory containing models in .safetensors format." 135 | 136 | print(f"[MetadataEditor] Copying metadata to models in {copy_dir}.") 137 | metadata = model_util.read_model_metadata(model_path, module) 138 | count = 0 139 | for entry in os.scandir(copy_dir): 140 | if entry.is_file(): 141 | path = os.path.realpath(os.path.join(copy_dir, entry.name)) 142 | if path != model_path and model_util.is_safetensors(path): 143 | if same_session_only: 144 | other_metadata = safetensors_hack.read_metadata(path) 145 | if missing_meta_only and other_metadata.get("ssmd_display_name", "").strip(): 146 | print(f"[MetadataEditor] Skipping {path} as it already has metadata") 147 | continue 148 | 149 | session_id = metadata.get("ss_session_id", None) 150 | other_session_id = other_metadata.get("ss_session_id", None) 151 | if session_id is None or other_session_id is None or session_id != other_session_id: 152 | continue 153 | 154 | updates = { 155 | "ssmd_cover_images": "[]", 156 | "ssmd_display_name": "", 157 | "ssmd_version": "", 158 | "ssmd_keywords": "", 159 | "ssmd_author": "", 160 | "ssmd_source": "", 161 | "ssmd_description": "", 162 | "ssmd_rating": "0", 163 | "ssmd_tags": "", 164 | } 165 | 166 | for k, v in metadata.items(): 167 | if k.startswith("ssmd_") and k != "ssmd_cover_images": 168 | updates[k] = v 169 | 170 | model_util.write_model_metadata(path, module, updates) 171 | count += 1 172 | 173 | print(f"[MetadataEditor] Updated {count} models in directory {copy_dir}.") 174 | return f"Updated {count} models in directory {copy_dir}." 175 | 176 | 177 | def load_cover_image(model_path, metadata): 178 | """ 179 | Loads a cover image either from embedded metadata or an image file with 180 | .preview.png/.png format 181 | """ 182 | cover_images = json.loads(metadata.get("ssmd_cover_images", "[]")) 183 | cover_image = None 184 | if len(cover_images) > 0: 185 | print("[MetadataEditor] Loading embedded cover image.") 186 | cover_image = decode_base64_to_pil(cover_images[0]) 187 | else: 188 | basename, ext = os.path.splitext(model_path) 189 | 190 | preview_paths = [f"{basename}.preview.png", f"{basename}.png"] 191 | 192 | for preview_path in preview_paths: 193 | if os.path.isfile(preview_path): 194 | print(f"[MetadataEditor] Loading webui preview image: {preview_path}") 195 | cover_image = Image.open(preview_path) 196 | 197 | return cover_image 198 | 199 | 200 | # Dummy value since gr.Dataframe cannot handle an empty list 201 | # https://github.com/gradio-app/gradio/issues/3182 202 | unknown_folders = ["(Unknown)", 0, 0, 0] 203 | 204 | 205 | def refresh_metadata(module, model_path): 206 | """ 207 | Reads metadata from the model on disk and updates all Gradio components 208 | """ 209 | if model_path == "None": 210 | return {}, None, "", "", "", "", "", 0, "", "", "", "", "", {}, [unknown_folders] 211 | 212 | if not os.path.isfile(model_path): 213 | return ( 214 | {"info": f"Model path not found: {model_path}"}, 215 | None, 216 | "", 217 | "", 218 | "", 219 | "", 220 | "", 221 | 0, 222 | "", 223 | "", 224 | "", 225 | "", 226 | "", 227 | {}, 228 | [unknown_folders], 229 | ) 230 | 231 | if os.path.splitext(model_path)[1] != ".safetensors": 232 | return ( 233 | {"info": "Model is not in .safetensors format."}, 234 | None, 235 | "", 236 | "", 237 | "", 238 | "", 239 | "", 240 | 0, 241 | "", 242 | "", 243 | "", 244 | "", 245 | "", 246 | {}, 247 | [unknown_folders], 248 | ) 249 | 250 | metadata = model_util.read_model_metadata(model_path, module) 251 | 252 | if metadata is None: 253 | training_params = {} 254 | metadata = {} 255 | else: 256 | training_params = {k: v for k, v in metadata.items() if k.startswith("ss_")} 257 | 258 | cover_image = load_cover_image(model_path, metadata) 259 | 260 | display_name = metadata.get("ssmd_display_name", "") 261 | author = metadata.get("ssmd_author", "") 262 | # version = metadata.get("ssmd_version", "") 263 | source = metadata.get("ssmd_source", "") 264 | keywords = metadata.get("ssmd_keywords", "") 265 | description = metadata.get("ssmd_description", "") 266 | rating = int(metadata.get("ssmd_rating", "0")) 267 | tags = metadata.get("ssmd_tags", "") 268 | model_hash = metadata.get("sshs_model_hash", model_util.cache("hashes").get(model_path, {}).get("model", "")) 269 | legacy_hash = metadata.get("sshs_legacy_hash", model_util.cache("hashes").get(model_path, {}).get("legacy", "")) 270 | 271 | top_tags = {} 272 | if "ss_tag_frequency" in training_params: 273 | tag_frequency = json.loads(training_params.pop("ss_tag_frequency")) 274 | count_max = 0 275 | for dir, frequencies in tag_frequency.items(): 276 | for tag, count in frequencies.items(): 277 | tag = tag.strip() 278 | existing = top_tags.get(tag, 0) 279 | top_tags[tag] = count + existing 280 | if len(top_tags) > 0: 281 | top_tags = dict(sorted(top_tags.items(), key=lambda x: x[1], reverse=True)) 282 | 283 | count_max = max(top_tags.values()) 284 | top_tags = {k: float(v / count_max) for k, v in top_tags.items()} 285 | 286 | dataset_folders = [] 287 | if "ss_dataset_dirs" in training_params: 288 | dataset_dirs = json.loads(training_params.pop("ss_dataset_dirs")) 289 | for dir, counts in dataset_dirs.items(): 290 | img_count = int(counts["img_count"]) 291 | n_repeats = int(counts["n_repeats"]) 292 | dataset_folders.append([dir, img_count, n_repeats, img_count * n_repeats]) 293 | if dataset_folders: 294 | dataset_folders.append( 295 | ["(Total)", sum(r[1] for r in dataset_folders), sum(r[2] for r in dataset_folders), sum(r[3] for r in dataset_folders)] 296 | ) 297 | else: 298 | dataset_folders.append(unknown_folders) 299 | 300 | return ( 301 | training_params, 302 | cover_image, 303 | display_name, 304 | author, 305 | source, 306 | keywords, 307 | description, 308 | rating, 309 | tags, 310 | model_hash, 311 | legacy_hash, 312 | model_path, 313 | os.path.dirname(model_path), 314 | top_tags, 315 | dataset_folders, 316 | ) 317 | 318 | 319 | def save_metadata(module, model_path, cover_image, display_name, author, source, keywords, description, rating, tags): 320 | """ 321 | Writes metadata from the Gradio components to the model file 322 | """ 323 | if model_path == "None": 324 | return "No model selected.", "", "" 325 | 326 | if not os.path.isfile(model_path): 327 | return f"file not found: {model_path}", "", "" 328 | 329 | if os.path.splitext(model_path)[1] != ".safetensors": 330 | return "Model is not in .safetensors format", "", "" 331 | 332 | metadata = safetensors_hack.read_metadata(model_path) 333 | model_hash = safetensors_hack.hash_file(model_path) 334 | legacy_hash = model_util.get_legacy_hash(metadata, model_path) 335 | 336 | # TODO: Support multiple images 337 | # Blocked on gradio not having a gallery upload option 338 | # https://github.com/gradio-app/gradio/issues/1379 339 | cover_images = [] 340 | if cover_image is not None: 341 | cover_images.append(encode_pil_to_base64(cover_image).decode("ascii")) 342 | 343 | # NOTE: User-specified metadata should NOT be prefixed with "ss_". This is 344 | # to maintain backwards compatibility with the old hashing method. "ss_" 345 | # should be used for training parameters that will never be manually 346 | # updated on the model. 347 | updates = { 348 | "ssmd_cover_images": json.dumps(cover_images), 349 | "ssmd_display_name": display_name, 350 | "ssmd_author": author, 351 | # "ssmd_version": version, 352 | "ssmd_source": source, 353 | "ssmd_keywords": keywords, 354 | "ssmd_description": description, 355 | "ssmd_rating": rating, 356 | "ssmd_tags": tags, 357 | "sshs_model_hash": model_hash, 358 | "sshs_legacy_hash": legacy_hash, 359 | } 360 | 361 | model_util.write_model_metadata(model_path, module, updates) 362 | if cover_image is None: 363 | delete_webui_model_preview_image(model_path) 364 | else: 365 | write_webui_model_preview_image(model_path, cover_image) 366 | 367 | model_name = os.path.basename(model_path) 368 | return f"Model saved: {model_name}", model_hash, legacy_hash 369 | 370 | 371 | model_name_filter = "" 372 | 373 | 374 | def get_filtered_model_paths(s): 375 | # newer Gradio seems to show None in the list? 376 | # if not s: 377 | # return ["None"] + list(model_util.lora_models.values()) 378 | # return ["None"] + [v for v in model_util.lora_models.values() if v and s in v.lower()] 379 | if not s: 380 | l = list(model_util.lora_models.values()) 381 | else: 382 | l = [v for v in model_util.lora_models.values() if v and s in v.lower()] 383 | l = [v for v in l if v] # remove None 384 | l = ["None"] + l 385 | return l 386 | 387 | def get_filtered_model_paths_global(): 388 | global model_name_filter 389 | return get_filtered_model_paths(model_name_filter) 390 | 391 | 392 | def setup_ui(addnet_paste_params): 393 | """ 394 | :dict addnet_paste_params: Dictionary of txt2img/img2img controls for each model weight slider, 395 | for sending module and model to them from the metadata editor 396 | """ 397 | can_edit = False 398 | 399 | with gr.Row().style(equal_height=False): 400 | # Lefthand column 401 | with gr.Column(variant="panel"): 402 | # Module and model selector 403 | with gr.Row(): 404 | model_filter = gr.Textbox("", label="Model path filter", placeholder="Filter models by path name") 405 | 406 | def update_model_filter(s): 407 | global model_name_filter 408 | model_name_filter = s.strip().lower() 409 | 410 | model_filter.change(update_model_filter, inputs=[model_filter], outputs=[]) 411 | with gr.Row(): 412 | module = gr.Dropdown( 413 | ["LoRA"], 414 | label="Network module", 415 | value="LoRA", 416 | interactive=True, 417 | elem_id="additional_networks_metadata_editor_module", 418 | ) 419 | model = gr.Dropdown( 420 | get_filtered_model_paths_global(), 421 | label="Model", 422 | value="None", 423 | interactive=True, 424 | elem_id="additional_networks_metadata_editor_model", 425 | ) 426 | modules.ui.create_refresh_button( 427 | model, model_util.update_models, lambda: {"choices": get_filtered_model_paths_global()}, "refresh_lora_models" 428 | ) 429 | 430 | def submit_model_filter(s): 431 | global model_name_filter 432 | model_name_filter = s 433 | paths = get_filtered_model_paths(s) 434 | return gr.Dropdown.update(choices=paths, value="None") 435 | 436 | model_filter.submit(submit_model_filter, inputs=[model_filter], outputs=[model]) 437 | 438 | # Model hashes and path 439 | with gr.Row(): 440 | model_hash = gr.Textbox("", label="Model hash", interactive=False) 441 | legacy_hash = gr.Textbox("", label="Legacy hash", interactive=False) 442 | with gr.Row(): 443 | model_path = gr.Textbox("", label="Model path", interactive=False) 444 | open_folder_button = ToolButton( 445 | value=folder_symbol, 446 | elem_id="hidden_element" if shared.cmd_opts.hide_ui_dir_config else "open_folder_metadata_editor", 447 | ) 448 | 449 | # Send to txt2img/img2img buttons 450 | for tabname in ["txt2img", "img2img"]: 451 | with gr.Row(): 452 | with gr.Box(): 453 | with gr.Row(): 454 | gr.HTML(f"Send to {tabname}:") 455 | for i in range(MAX_MODEL_COUNT): 456 | send_to_button = ToolButton( 457 | value=keycap_symbols[i], elem_id=f"additional_networks_send_to_{tabname}_{i}" 458 | ) 459 | send_to_button.click( 460 | fn=lambda modu, mod: (modu, model_util.find_closest_lora_model_name(mod) or "None"), 461 | inputs=[module, model], 462 | outputs=[addnet_paste_params[tabname][i]["module"], addnet_paste_params[tabname][i]["model"]], 463 | ) 464 | send_to_button.click(fn=None, _js=f"addnet_switch_to_{tabname}", inputs=None, outputs=None) 465 | 466 | # "Copy metadata to other models" panel 467 | with gr.Row(): 468 | with gr.Column(): 469 | gr.HTML(value="Copy metadata to other models in directory") 470 | copy_metadata_dir = gr.Textbox( 471 | "", 472 | label="Containing directory", 473 | placeholder="All models in this directory will receive the selected model's metadata", 474 | ) 475 | with gr.Row(): 476 | copy_same_session = gr.Checkbox(True, label="Only copy to models with same session ID") 477 | copy_no_metadata = gr.Checkbox(True, label="Only copy to models with no metadata") 478 | copy_metadata_button = gr.Button("Copy Metadata", variant="primary") 479 | 480 | # Center column, metadata viewer/editor 481 | with gr.Column(): 482 | with gr.Row(): 483 | display_name = gr.Textbox(value="", label="Name", placeholder="Display name for this model", interactive=can_edit) 484 | author = gr.Textbox(value="", label="Author", placeholder="Author of this model", interactive=can_edit) 485 | with gr.Row(): 486 | keywords = gr.Textbox( 487 | value="", label="Keywords", placeholder="Activation keywords, comma-separated", interactive=can_edit 488 | ) 489 | with gr.Row(): 490 | description = gr.Textbox( 491 | value="", 492 | label="Description", 493 | placeholder="Model description/readme/notes/instructions", 494 | lines=15, 495 | interactive=can_edit, 496 | ) 497 | with gr.Row(): 498 | source = gr.Textbox( 499 | value="", label="Source", placeholder="Source URL where this model could be found", interactive=can_edit 500 | ) 501 | with gr.Row(): 502 | rating = gr.Slider(minimum=0, maximum=10, step=1, label="Rating", value=0, interactive=can_edit) 503 | tags = gr.Textbox( 504 | value="", 505 | label="Tags", 506 | placeholder='Comma-separated list of tags ("artist, style, character, 2d, 3d...")', 507 | lines=2, 508 | interactive=can_edit, 509 | ) 510 | with gr.Row(): 511 | editing_enabled = gr.Checkbox(label="Editing Enabled", value=can_edit) 512 | with gr.Row(): 513 | save_metadata_button = gr.Button("Save Metadata", variant="primary", interactive=can_edit) 514 | with gr.Row(): 515 | save_output = gr.HTML("") 516 | 517 | # Righthand column, cover image and training parameters view 518 | with gr.Column(): 519 | # Cover image 520 | with gr.Row(): 521 | cover_image = gr.Image( 522 | label="Cover image", 523 | elem_id="additional_networks_cover_image", 524 | source="upload", 525 | interactive=can_edit, 526 | type="pil", 527 | image_mode="RGBA", 528 | ).style(height=480) 529 | 530 | # Image parameters 531 | with gr.Accordion("Image Parameters", open=False): 532 | with gr.Row(): 533 | info2 = gr.HTML() 534 | with gr.Row(): 535 | try: 536 | send_to_buttons = parameters_copypaste.create_buttons(["txt2img", "img2img", "inpaint", "extras"]) 537 | except: 538 | pass 539 | 540 | # Training info, below cover image 541 | with gr.Accordion("Training info", open=False): 542 | # Top tags used 543 | with gr.Row(): 544 | max_top_tags = int(shared.opts.data.get("additional_networks_max_top_tags", 20)) 545 | most_frequent_tags = gr.Label(value={}, label="Most frequent tags in captions", num_top_classes=max_top_tags) 546 | 547 | # Dataset folders 548 | with gr.Row(): 549 | max_dataset_folders = int(shared.opts.data.get("additional_networks_max_dataset_folders", 20)) 550 | dataset_folders = gr.Dataframe( 551 | headers=["Name", "Image Count", "Repeats", "Total Images"], 552 | datatype=["str", "number", "number", "number"], 553 | label="Dataset folder structure", 554 | max_rows=max_dataset_folders, 555 | col_count=(4, "fixed"), 556 | ) 557 | 558 | # Training Parameters 559 | with gr.Row(): 560 | metadata_view = gr.JSON(value={}, label="Training parameters") 561 | 562 | # Hidden/internal 563 | with gr.Row(visible=False): 564 | info1 = gr.HTML() 565 | img_file_info = gr.Textbox(label="Generate Info", interactive=False, lines=6) 566 | 567 | open_folder_button.click(fn=lambda p: open_folder(os.path.dirname(p)), inputs=[model_path], outputs=[]) 568 | copy_metadata_button.click( 569 | fn=copy_metadata_to_all, 570 | inputs=[module, model, copy_metadata_dir, copy_same_session, copy_no_metadata, cover_image], 571 | outputs=[save_output], 572 | ) 573 | 574 | def update_editing(enabled): 575 | """ 576 | Enable/disable components based on "Editing Enabled" status 577 | """ 578 | updates = [gr.Textbox.update(interactive=enabled)] * 6 579 | updates.append(gr.Image.update(interactive=enabled)) 580 | updates.append(gr.Slider.update(interactive=enabled)) 581 | updates.append(gr.Button.update(interactive=enabled)) 582 | return updates 583 | 584 | editing_enabled.change( 585 | fn=update_editing, 586 | inputs=[editing_enabled], 587 | outputs=[display_name, author, source, keywords, description, tags, cover_image, rating, save_metadata_button], 588 | ) 589 | 590 | cover_image.change(fn=modules.extras.run_pnginfo, inputs=[cover_image], outputs=[info1, img_file_info, info2]) 591 | 592 | try: 593 | parameters_copypaste.bind_buttons(send_to_buttons, cover_image, img_file_info) 594 | except: 595 | pass 596 | 597 | model.change( 598 | refresh_metadata, 599 | inputs=[module, model], 600 | outputs=[ 601 | metadata_view, 602 | cover_image, 603 | display_name, 604 | author, 605 | source, 606 | keywords, 607 | description, 608 | rating, 609 | tags, 610 | model_hash, 611 | legacy_hash, 612 | model_path, 613 | copy_metadata_dir, 614 | most_frequent_tags, 615 | dataset_folders, 616 | ], 617 | ) 618 | save_metadata_button.click( 619 | save_metadata, 620 | inputs=[module, model, cover_image, display_name, author, source, keywords, description, rating, tags], 621 | outputs=[save_output, model_hash, legacy_hash], 622 | ) 623 | -------------------------------------------------------------------------------- /scripts/model_util.py: -------------------------------------------------------------------------------- 1 | import os 2 | import os.path 3 | import re 4 | import shutil 5 | import json 6 | import stat 7 | import tqdm 8 | from collections import OrderedDict 9 | from multiprocessing.pool import ThreadPool as Pool 10 | 11 | from modules import shared, sd_models, hashes 12 | from scripts import safetensors_hack, model_util, util 13 | import modules.scripts as scripts 14 | 15 | 16 | # MAX_MODEL_COUNT = shared.cmd_opts.addnet_max_model_count or 5 17 | MAX_MODEL_COUNT = shared.cmd_opts.addnet_max_model_count if hasattr(shared.cmd_opts, "addnet_max_model_count") else 5 18 | LORA_MODEL_EXTS = [".pt", ".ckpt", ".safetensors"] 19 | re_legacy_hash = re.compile("\(([0-9a-f]{8})\)$") # matches 8-character hashes, new hash has 12 characters 20 | lora_models = {} # "My_Lora(abcdef123456)" -> "C:/path/to/model.safetensors" 21 | lora_model_names = {} # "my_lora" -> "My_Lora(My_Lora(abcdef123456)" 22 | legacy_model_names = {} 23 | lora_models_dir = os.path.join(scripts.basedir(), "models/lora") 24 | os.makedirs(lora_models_dir, exist_ok=True) 25 | 26 | 27 | def is_safetensors(filename): 28 | return os.path.splitext(filename)[1] == ".safetensors" 29 | 30 | 31 | def read_model_metadata(model_path, module): 32 | if model_path.startswith('"') and model_path.endswith('"'): # trim '"' at start/end 33 | model_path = model_path[1:-1] 34 | if not os.path.exists(model_path): 35 | return None 36 | 37 | metadata = None 38 | if module == "LoRA": 39 | if os.path.splitext(model_path)[1] == ".safetensors": 40 | metadata = safetensors_hack.read_metadata(model_path) 41 | 42 | return metadata 43 | 44 | 45 | def write_model_metadata(model_path, module, updates): 46 | if model_path.startswith('"') and model_path.endswith('"'): # trim '"' at start/end 47 | model_path = model_path[1:-1] 48 | if not os.path.exists(model_path): 49 | return None 50 | 51 | from safetensors.torch import save_file 52 | 53 | back_up = shared.opts.data.get("additional_networks_back_up_model_when_saving", True) 54 | if back_up: 55 | backup_path = model_path + ".backup" 56 | if not os.path.exists(backup_path): 57 | print(f"[MetadataEditor] Backing up current model to {backup_path}") 58 | shutil.copyfile(model_path, backup_path) 59 | 60 | metadata = None 61 | tensors = {} 62 | if module == "LoRA": 63 | if os.path.splitext(model_path)[1] == ".safetensors": 64 | tensors, metadata = safetensors_hack.load_file(model_path, "cpu") 65 | 66 | for k, v in updates.items(): 67 | metadata[k] = str(v) 68 | 69 | save_file(tensors, model_path, metadata) 70 | print(f"[MetadataEditor] Model saved: {model_path}") 71 | 72 | 73 | def get_model_list(module, model, model_dir, sort_by): 74 | if model_dir == "": 75 | # Get list of models with same folder as this one 76 | model_path = lora_models.get(model, None) 77 | if model_path is None: 78 | return [] 79 | model_dir = os.path.dirname(model_path) 80 | 81 | if not os.path.isdir(model_dir): 82 | return [] 83 | 84 | found, _ = get_all_models([model_dir], sort_by, "") 85 | return list(found.keys()) # convert dict_keys to list 86 | 87 | 88 | def traverse_all_files(curr_path, model_list): 89 | f_list = [(os.path.join(curr_path, entry.name), entry.stat()) for entry in os.scandir(curr_path)] 90 | for f_info in f_list: 91 | fname, fstat = f_info 92 | if os.path.splitext(fname)[1] in LORA_MODEL_EXTS: 93 | model_list.append(f_info) 94 | elif stat.S_ISDIR(fstat.st_mode): 95 | model_list = traverse_all_files(fname, model_list) 96 | return model_list 97 | 98 | 99 | def get_model_hash(metadata, filename): 100 | if metadata is None: 101 | return hashes.calculate_sha256(filename) 102 | 103 | if "sshs_model_hash" in metadata: 104 | return metadata["sshs_model_hash"] 105 | 106 | return safetensors_hack.hash_file(filename) 107 | 108 | 109 | def get_legacy_hash(metadata, filename): 110 | if metadata is None: 111 | return sd_models.model_hash(filename) 112 | 113 | if "sshs_legacy_hash" in metadata: 114 | return metadata["sshs_legacy_hash"] 115 | 116 | return safetensors_hack.legacy_hash_file(filename) 117 | 118 | 119 | import filelock 120 | 121 | cache_filename = os.path.join(scripts.basedir(), "hashes.json") 122 | cache_data = None 123 | 124 | 125 | def cache(subsection): 126 | global cache_data 127 | 128 | if cache_data is None: 129 | with filelock.FileLock(cache_filename + ".lock"): 130 | if not os.path.isfile(cache_filename): 131 | cache_data = {} 132 | else: 133 | with open(cache_filename, "r", encoding="utf8") as file: 134 | cache_data = json.load(file) 135 | 136 | s = cache_data.get(subsection, {}) 137 | cache_data[subsection] = s 138 | 139 | return s 140 | 141 | 142 | def dump_cache(): 143 | with filelock.FileLock(cache_filename + ".lock"): 144 | with open(cache_filename, "w", encoding="utf8") as file: 145 | json.dump(cache_data, file, indent=4) 146 | 147 | 148 | def get_model_rating(filename): 149 | if not model_util.is_safetensors(filename): 150 | return 0 151 | 152 | metadata = safetensors_hack.read_metadata(filename) 153 | return int(metadata.get("ssmd_rating", "0")) 154 | 155 | 156 | def has_user_metadata(filename): 157 | if not model_util.is_safetensors(filename): 158 | return False 159 | 160 | metadata = safetensors_hack.read_metadata(filename) 161 | return any(k.startswith("ssmd_") for k in metadata.keys()) 162 | 163 | 164 | def hash_model_file(finfo): 165 | filename = finfo[0] 166 | stat = finfo[1] 167 | name = os.path.splitext(os.path.basename(filename))[0] 168 | 169 | # Prevent a hypothetical "None.pt" from being listed. 170 | if name != "None": 171 | metadata = None 172 | 173 | cached = cache("hashes").get(filename, None) 174 | if cached is None or stat.st_mtime != cached["mtime"]: 175 | if metadata is None and model_util.is_safetensors(filename): 176 | try: 177 | metadata = safetensors_hack.read_metadata(filename) 178 | except Exception as ex: 179 | return {"error": ex, "filename": filename} 180 | model_hash = get_model_hash(metadata, filename) 181 | legacy_hash = get_legacy_hash(metadata, filename) 182 | else: 183 | model_hash = cached["model"] 184 | legacy_hash = cached["legacy"] 185 | 186 | return {"model": model_hash, "legacy": legacy_hash, "fileinfo": finfo} 187 | 188 | 189 | def get_all_models(paths, sort_by, filter_by): 190 | fileinfos = [] 191 | for path in paths: 192 | if os.path.isdir(path): 193 | fileinfos += traverse_all_files(path, []) 194 | 195 | show_only_safetensors = shared.opts.data.get("additional_networks_show_only_safetensors", False) 196 | show_only_missing_meta = shared.opts.data.get("additional_networks_show_only_models_with_metadata", "disabled") 197 | 198 | if show_only_safetensors: 199 | fileinfos = [x for x in fileinfos if is_safetensors(x[0])] 200 | 201 | if show_only_missing_meta == "has metadata": 202 | fileinfos = [x for x in fileinfos if has_user_metadata(x[0])] 203 | elif show_only_missing_meta == "missing metadata": 204 | fileinfos = [x for x in fileinfos if not has_user_metadata(x[0])] 205 | 206 | print("[AddNet] Updating model hashes...") 207 | data = [] 208 | thread_count = max(1, int(shared.opts.data.get("additional_networks_hash_thread_count", 1))) 209 | p = Pool(processes=thread_count) 210 | with tqdm.tqdm(total=len(fileinfos)) as pbar: 211 | for res in p.imap_unordered(hash_model_file, fileinfos): 212 | pbar.update() 213 | if "error" in res: 214 | print(f"Failed to read model file {res['filename']}: {res['error']}") 215 | else: 216 | data.append(res) 217 | p.close() 218 | 219 | cache_hashes = cache("hashes") 220 | 221 | res = OrderedDict() 222 | res_legacy = OrderedDict() 223 | filter_by = filter_by.strip(" ") 224 | if len(filter_by) != 0: 225 | data = [x for x in data if filter_by.lower() in os.path.basename(x["fileinfo"][0]).lower()] 226 | if sort_by == "name": 227 | data = sorted(data, key=lambda x: os.path.basename(x["fileinfo"][0])) 228 | elif sort_by == "date": 229 | data = sorted(data, key=lambda x: -x["fileinfo"][1].st_mtime) 230 | elif sort_by == "path name": 231 | data = sorted(data, key=lambda x: x["fileinfo"][0]) 232 | elif sort_by == "rating": 233 | data = sorted(data, key=lambda x: get_model_rating(x["fileinfo"][0]), reverse=True) 234 | elif sort_by == "has user metadata": 235 | data = sorted( 236 | data, key=lambda x: os.path.basename(x["fileinfo"][0]) if has_user_metadata(x["fileinfo"][0]) else "", reverse=True 237 | ) 238 | 239 | reverse = shared.opts.data.get("additional_networks_reverse_sort_order", False) 240 | if reverse: 241 | data = reversed(data) 242 | 243 | for result in data: 244 | finfo = result["fileinfo"] 245 | filename = finfo[0] 246 | stat = finfo[1] 247 | model_hash = result["model"] 248 | legacy_hash = result["legacy"] 249 | 250 | name = os.path.splitext(os.path.basename(filename))[0] 251 | 252 | # Commas in the model name will mess up infotext restoration since the 253 | # infotext is delimited by commas 254 | name = name.replace(",", "_") 255 | 256 | # Prevent a hypothetical "None.pt" from being listed. 257 | if name != "None": 258 | full_name = name + f"({model_hash[0:12]})" 259 | res[full_name] = filename 260 | res_legacy[legacy_hash] = full_name 261 | cache_hashes[filename] = {"model": model_hash, "legacy": legacy_hash, "mtime": stat.st_mtime} 262 | 263 | return res, res_legacy 264 | 265 | 266 | def find_closest_lora_model_name(search: str): 267 | if not search or search == "None": 268 | return None 269 | 270 | # Match name and hash, case-sensitive 271 | # "MyModel-epoch00002(abcdef123456)" 272 | if search in lora_models: 273 | return search 274 | 275 | # Match model path, case-sensitive (from metadata editor) 276 | # "C:/path/to/mymodel-epoch00002.safetensors" 277 | if os.path.isfile(search): 278 | import json 279 | 280 | find = os.path.normpath(search) 281 | value = next((k for k in lora_models.keys() if lora_models[k] == find), None) 282 | if value: 283 | return value 284 | 285 | search = search.lower() 286 | 287 | # Match full name, case-insensitive 288 | # "mymodel-epoch00002" 289 | if search in lora_model_names: 290 | return lora_model_names.get(search) 291 | 292 | # Match legacy hash (8 characters) 293 | # "MyModel(abcd1234)" 294 | result = re_legacy_hash.search(search) 295 | if result is not None: 296 | model_hash = result.group(1) 297 | if model_hash in legacy_model_names: 298 | new_model_name = legacy_model_names[model_hash] 299 | return new_model_name 300 | 301 | # Use any model with the search term as the prefix, case-insensitive, sorted 302 | # by name length 303 | # "mymodel" 304 | applicable = [name for name in lora_model_names.keys() if search in name.lower()] 305 | if not applicable: 306 | return None 307 | applicable = sorted(applicable, key=lambda name: len(name)) 308 | return lora_model_names[applicable[0]] 309 | 310 | 311 | def update_models(): 312 | global lora_models, lora_model_names, legacy_model_names 313 | paths = [lora_models_dir] 314 | extra_lora_paths = util.split_path_list(shared.opts.data.get("additional_networks_extra_lora_path", "")) 315 | for path in extra_lora_paths: 316 | path = path.lstrip() 317 | if os.path.isdir(path): 318 | paths.append(path) 319 | 320 | sort_by = shared.opts.data.get("additional_networks_sort_models_by", "name") 321 | filter_by = shared.opts.data.get("additional_networks_model_name_filter", "") 322 | res, res_legacy = get_all_models(paths, sort_by, filter_by) 323 | 324 | lora_models.clear() 325 | lora_models["None"] = None 326 | lora_models.update(res) 327 | 328 | for name_and_hash, filename in lora_models.items(): 329 | if filename == None: 330 | continue 331 | name = os.path.splitext(os.path.basename(filename))[0].lower() 332 | lora_model_names[name] = name_and_hash 333 | 334 | legacy_model_names = res_legacy 335 | dump_cache() 336 | 337 | 338 | update_models() 339 | -------------------------------------------------------------------------------- /scripts/safetensors_hack.py: -------------------------------------------------------------------------------- 1 | import io 2 | import os 3 | import mmap 4 | import torch 5 | import json 6 | import hashlib 7 | import safetensors 8 | import safetensors.torch 9 | 10 | from modules import sd_models 11 | 12 | # PyTorch 1.13 and later have _UntypedStorage renamed to UntypedStorage 13 | UntypedStorage = torch.storage.UntypedStorage if hasattr(torch.storage, 'UntypedStorage') else torch.storage._UntypedStorage 14 | 15 | def read_metadata(filename): 16 | """Reads the JSON metadata from a .safetensors file""" 17 | with open(filename, mode="r", encoding="utf8") as file_obj: 18 | with mmap.mmap(file_obj.fileno(), length=0, access=mmap.ACCESS_READ) as m: 19 | header = m.read(8) 20 | n = int.from_bytes(header, "little") 21 | metadata_bytes = m.read(n) 22 | metadata = json.loads(metadata_bytes) 23 | 24 | return metadata.get("__metadata__", {}) 25 | 26 | 27 | def load_file(filename, device): 28 | """"Loads a .safetensors file without memory mapping that locks the model file. 29 | Works around safetensors issue: https://github.com/huggingface/safetensors/issues/164""" 30 | with open(filename, mode="r", encoding="utf8") as file_obj: 31 | with mmap.mmap(file_obj.fileno(), length=0, access=mmap.ACCESS_READ) as m: 32 | header = m.read(8) 33 | n = int.from_bytes(header, "little") 34 | metadata_bytes = m.read(n) 35 | metadata = json.loads(metadata_bytes) 36 | 37 | size = os.stat(filename).st_size 38 | storage = UntypedStorage.from_file(filename, False, size) 39 | offset = n + 8 40 | md = metadata.get("__metadata__", {}) 41 | return {name: create_tensor(storage, info, offset) for name, info in metadata.items() if name != "__metadata__"}, md 42 | 43 | 44 | def hash_file(filename): 45 | """Hashes a .safetensors file using the new hashing method. 46 | Only hashes the weights of the model.""" 47 | hash_sha256 = hashlib.sha256() 48 | blksize = 1024 * 1024 49 | 50 | with open(filename, mode="r", encoding="utf8") as file_obj: 51 | with mmap.mmap(file_obj.fileno(), length=0, access=mmap.ACCESS_READ) as m: 52 | header = m.read(8) 53 | n = int.from_bytes(header, "little") 54 | 55 | with open(filename, mode="rb") as file_obj: 56 | offset = n + 8 57 | file_obj.seek(offset) 58 | for chunk in iter(lambda: file_obj.read(blksize), b""): 59 | hash_sha256.update(chunk) 60 | 61 | return hash_sha256.hexdigest() 62 | 63 | 64 | def legacy_hash_file(filename): 65 | """Hashes a model file using the legacy `sd_models.model_hash()` method.""" 66 | hash_sha256 = hashlib.sha256() 67 | 68 | metadata = read_metadata(filename) 69 | 70 | # For compatibility with legacy models: This replicates the behavior of 71 | # sd_models.model_hash as if there were no user-specified metadata in the 72 | # .safetensors file. That leaves the training parameters, which are 73 | # immutable. It is important the hash does not include the embedded user 74 | # metadata as that would mean the hash could change every time the user 75 | # updates the name/description/etc. The new hashing method fixes this 76 | # problem by only hashing the region of the file containing the tensors. 77 | if any(not k.startswith("ss_") for k in metadata): 78 | # Strip the user metadata, re-serialize the file as if it were freshly 79 | # created from sd-scripts, and hash that with model_hash's behavior. 80 | tensors, metadata = load_file(filename, "cpu") 81 | metadata = {k: v for k, v in metadata.items() if k.startswith("ss_")} 82 | model_bytes = safetensors.torch.save(tensors, metadata) 83 | 84 | hash_sha256.update(model_bytes[0x100000:0x110000]) 85 | return hash_sha256.hexdigest()[0:8] 86 | else: 87 | # This should work fine with model_hash since when the legacy hashing 88 | # method was being used the user metadata system hadn't been implemented 89 | # yet. 90 | return sd_models.model_hash(filename) 91 | 92 | 93 | DTYPES = { 94 | "F64": torch.float64, 95 | "F32": torch.float32, 96 | "F16": torch.float16, 97 | "BF16": torch.bfloat16, 98 | "I64": torch.int64, 99 | # "U64": torch.uint64, 100 | "I32": torch.int32, 101 | # "U32": torch.uint32, 102 | "I16": torch.int16, 103 | # "U16": torch.uint16, 104 | "I8": torch.int8, 105 | "U8": torch.uint8, 106 | "BOOL": torch.bool 107 | } 108 | 109 | 110 | def create_tensor(storage, info, offset): 111 | """Creates a tensor without holding on to an open handle to the parent model 112 | file.""" 113 | dtype = DTYPES[info["dtype"]] 114 | shape = info["shape"] 115 | start, stop = info["data_offsets"] 116 | return torch.asarray(storage[start + offset : stop + offset], dtype=torch.uint8).view(dtype=dtype).reshape(shape).clone().detach() 117 | -------------------------------------------------------------------------------- /scripts/util.py: -------------------------------------------------------------------------------- 1 | import csv 2 | from io import StringIO 3 | from typing import List 4 | 5 | def split_path_list(path_list: str) -> List[str]: 6 | pl = [] 7 | with StringIO() as f: 8 | f.write(path_list) 9 | f.seek(0) 10 | for r in csv.reader(f): 11 | pl += r 12 | return pl 13 | -------------------------------------------------------------------------------- /style.css: -------------------------------------------------------------------------------- 1 | #additional_networks_cover_image, 2 | #additional_networks_cover_image > .h-60, 3 | #additional_networks_cover_image > .h-60 > div, 4 | #additional_networks_cover_image > .h-60 > div > img 5 | { 6 | height: 480px !important; 7 | max-height: 480px !important; 8 | min-height: 480px !important; 9 | } 10 | --------------------------------------------------------------------------------