├── LICENSE ├── README.md ├── __init__.py ├── evaluation.py ├── examples ├── sdxl_kanagawa_512x512.png └── upscale.jpg ├── latent_resizer.py ├── latent_resizer_train.py ├── nn_upscale.py ├── sd15_resizer.pt └── sdxl_resizer.pt /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # ComfyUI Neural network latent upscale custom node 2 | 3 | ![Example 1](./examples/sdxl_kanagawa_512x512.png) 4 | 5 | This repository includes a custom node for 6 | [ComfyUI](https://github.com/comfyanonymous/ComfyUI) for upscaling the latents 7 | quickly using a small neural network without needing to decode and encode with 8 | VAE. The node can be found in "Add Node -> latent -> NNLatentUpscale". 9 | 10 | This node is meant to be used in a workflow where the initial image is 11 | generated in lower resolution, the latent is upscaled and the upscaled latent is 12 | fed back into the stable diffusion u-net for low noise diffusion pass (high-res 13 | fix). 14 | 15 | Compared to VAE decode -> upscale -> encode, the neural net latent upscale is 16 | about 20 - 50 times faster depending on the image resolution with minimal 17 | quality loss. Compared to direct linear interpolation of the latent the neural 18 | net upscale is slower but has much better quality. Direct latent interpolation 19 | usually has very large artifacts. 20 | 21 | ## Installation 22 | 23 | Clone this repository in ComfyUI `custom_nodes` directory with: `git clone https://github.com/Ttl/ComfyUi_NNLatentUpscale.git`. 24 | 25 | ## Evaluation 26 | 27 | ![Example 2](./examples/upscale.jpg) 28 | 29 | Dataset: [COCO 2017](https://cocodataset.org.org) validation images center 30 | cropped to 256x256 resolution. The comparison image is linear upscale of the 31 | input image. All tests are done with fp32 precision and batch size 4. 32 | 33 | 34 | VAE Upscale: VAE decode -> Linear interpolation -> Encode. 35 | 36 | NN Upscale: Neural network upscale (This repository). 37 | 38 | Latent Upscale: Linear interpolation of latent. 39 | 40 | SDXL, 2x upscale: 41 | 42 | | | MSE ↓ | LPIPS ↓ | PSNR ↑ | Time (ms) ↓ | 43 | |----------------------|--------|---------|--------|-------------| 44 | | VAE Upscale | 0.009 | 0.22 | 26.9 | 832 | 45 | | NN Upscale | 0.010 | 0.28 | 26.3 | 36 | 46 | | Latent Upscale | 0.047 | 0.65 | 19.5 | 0.1 | 47 | 48 | SDXL, 1.5x upscale: 49 | 50 | | | MSE ↓ | LPIPS ↓ | PSNR ↑ | Time (ms) ↓ | 51 | |----------------------|--------|---------|--------|-------------| 52 | | VAE Upscale | 0.009 | 0.20 | 26.9 | 583 | 53 | | NN Upscale | 0.010 | 0.26 | 26.3 | 19 | 54 | | Latent Upscale | 0.038 | 0.58 | 20.4 | 0.1 | 55 | 56 | SD 1.5, 2x upscale: 57 | 58 | | | MSE ↓ | LPIPS ↓ | PSNR ↑ | Time (ms) ↓ | 59 | |----------------------|-------|---------|--------|-------------| 60 | | VAE Upscale | 0.009 | 0.21 | 26.7 | 822 | 61 | | NN Upscale | 0.008 | 0.24 | 27.0 | 36 | 62 | | Latent Upscale | 0.033 | 0.61 | 20.9 | 0.1 | 63 | 64 | SD 1.5, 1.5x upscale: 65 | 66 | | | MSE ↓ | LPIPS ↓ | PSNR ↑ | Time (ms) ↓ | 67 | |----------------------|-------|---------|--------|-------------| 68 | | VAE Upscale | 0.010 | 0.18 | 26.5 | 594 | 69 | | NN Upscale | 0.009 | 0.21 | 26.9 | 20 | 70 | | Latent Upscale | 0.031 | 0.52 | 21.3 | 0.1 | 71 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | from .nn_upscale import NNLatentUpscale 2 | 3 | NODE_CLASS_MAPPINGS = { 4 | "NNLatentUpscale": NNLatentUpscale 5 | } 6 | 7 | NODE_DISPLAY_NAME_MAPPINGS = { 8 | "NNlLatentUpscale": "NN Latent Upscale" 9 | } 10 | -------------------------------------------------------------------------------- /evaluation.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | from latent_resizer import LatentResizer 3 | import argparse 4 | from diffusers import AutoencoderKL 5 | import lpips 6 | import torch 7 | from torchvision import transforms 8 | from pathlib import Path 9 | from PIL import Image 10 | import numpy as np 11 | from tqdm import tqdm 12 | from pytorch_msssim import ssim 13 | 14 | 15 | def psnr(x, ref, maxg=2): 16 | mse = torch.mean(torch.square(x - ref)) 17 | return 20 * torch.log10(maxg / torch.sqrt(mse)) 18 | 19 | 20 | class ImageDataset(torch.utils.data.Dataset): 21 | def __init__(self, path, size): 22 | self.path = Path(path) 23 | if not self.path.exists(): 24 | raise ValueError("Dataset path does not exist") 25 | self.images = list(self.path.iterdir()) 26 | self.num_images = len(self.images) 27 | 28 | self.image_transforms = transforms.Compose( 29 | [ 30 | transforms.Resize( 31 | size, interpolation=transforms.InterpolationMode.BILINEAR 32 | ), 33 | transforms.CenterCrop(size), 34 | transforms.ToTensor(), 35 | transforms.Normalize([0.5], [0.5]), 36 | ] 37 | ) 38 | 39 | def __len__(self): 40 | return self.num_images 41 | 42 | def __getitem__(self, index): 43 | example = {} 44 | image = Image.open(self.images[index % self.num_images]).convert("RGB") 45 | image = self.image_transforms(image) 46 | return image 47 | 48 | 49 | def collate_fn(images): 50 | images = torch.stack(images) 51 | images = images.to(memory_format=torch.contiguous_format).float() 52 | return images 53 | 54 | 55 | if __name__ == "__main__": 56 | parser = argparse.ArgumentParser(description="Latent resizer evaluation") 57 | parser.add_argument( 58 | "--test_path", 59 | required=True, 60 | type=str, 61 | help="Test images", 62 | ) 63 | parser.add_argument( 64 | "--vae_path", 65 | required=True, 66 | type=str, 67 | help="VAE path", 68 | ) 69 | parser.add_argument( 70 | "--resizer_path", 71 | required=True, 72 | type=str, 73 | help="Resizer weight path", 74 | ) 75 | parser.add_argument( 76 | "--device", 77 | required=False, 78 | type=str, 79 | default="cuda", 80 | help="Torch device", 81 | ) 82 | parser.add_argument( 83 | "--fp16", 84 | action="store_true", 85 | help="Use fp16 precision", 86 | ) 87 | parser.add_argument( 88 | "--num_workers", 89 | default=4, 90 | type=int, 91 | help="Dataloader workers", 92 | ) 93 | parser.add_argument( 94 | "--batch_size", 95 | default=8, 96 | type=int, 97 | help="Batch size", 98 | ) 99 | parser.add_argument( 100 | "--resolution", 101 | default=256, 102 | type=int, 103 | help="Image resolution", 104 | ) 105 | parser.add_argument( 106 | "--scale", 107 | default=2.0, 108 | type=float, 109 | required=True, 110 | help="Resize scale", 111 | ) 112 | parser.add_argument( 113 | "--resizer_only", 114 | action="store_true", 115 | help="Only evaluate resizer", 116 | ) 117 | 118 | args = parser.parse_args() 119 | device = torch.device(args.device) 120 | scale_factor = 0.13025 121 | 122 | if args.fp16: 123 | dtype = torch.float16 124 | else: 125 | dtype = torch.float32 126 | 127 | vae = AutoencoderKL.from_single_file(args.vae_path).to(device, dtype=dtype) 128 | vae.eval() 129 | 130 | resizer = LatentResizer.load_model(args.resizer_path, device, dtype) 131 | 132 | # LPIPS is always in float32 because of nans in float16 133 | lpips_fn = lpips.LPIPS(net="vgg").to(device=device, dtype=torch.float32) 134 | 135 | dataset = ImageDataset(args.test_path, args.resolution) 136 | dataloader = torch.utils.data.DataLoader( 137 | dataset, batch_size=args.batch_size, num_workers=args.num_workers 138 | ) 139 | 140 | elapsed_vae_list = [] 141 | elapsed_resizer_list = [] 142 | elapsed_latent_list = [] 143 | 144 | mse_vae_list = [] 145 | mse_resizer_list = [] 146 | mse_latent_list = [] 147 | 148 | lpips_vae_list = [] 149 | lpips_resizer_list = [] 150 | lpips_latent_list = [] 151 | 152 | psnr_vae_list = [] 153 | psnr_resizer_list = [] 154 | psnr_latent_list = [] 155 | 156 | ssim_vae_list = [] 157 | ssim_resizer_list = [] 158 | ssim_latent_list = [] 159 | 160 | try: 161 | with torch.inference_mode(): 162 | for images in tqdm(dataloader): 163 | images = images.to(device=device, dtype=dtype) 164 | images_upscaled = torch.nn.functional.interpolate( 165 | images, scale_factor=args.scale, mode="bilinear" 166 | ) 167 | latents = vae.encode(images).latent_dist.sample() 168 | del images 169 | 170 | # Resizer 171 | start_resizer = torch.cuda.Event(enable_timing=True) 172 | end_resizer = torch.cuda.Event(enable_timing=True) 173 | start_resizer.record() 174 | resized = ( 175 | resizer(scale_factor * latents, scale=args.scale) / scale_factor 176 | ) 177 | end_resizer.record() 178 | torch.cuda.synchronize() 179 | resizer_elapsed = start_resizer.elapsed_time(end_resizer) 180 | 181 | decoded_resized = vae.decode(resized)[0] 182 | del resized 183 | 184 | elapsed_resizer_list.append(resizer_elapsed) 185 | 186 | mse_resizer = torch.nn.functional.mse_loss( 187 | decoded_resized, images_upscaled 188 | ) 189 | mse_resizer_list.extend(mse_resizer.cpu().numpy().flatten()) 190 | 191 | lpips_resizer = lpips_fn( 192 | decoded_resized.float(), images_upscaled.float() 193 | ) 194 | 195 | lpips_resizer_list.extend(lpips_resizer.cpu().numpy().flatten()) 196 | 197 | psnr_resized = psnr(decoded_resized, images_upscaled) 198 | psnr_resizer_list.extend(psnr_resized.cpu().numpy().flatten()) 199 | 200 | ssim_resized = ssim( 201 | 0.5 * (decoded_resized + 1), 202 | 0.5 * (images_upscaled + 1), 203 | data_range=1, 204 | size_average=True, 205 | ) 206 | ssim_resizer_list.append(ssim_resized.cpu()) 207 | 208 | if not args.resizer_only: 209 | start_vae = torch.cuda.Event(enable_timing=True) 210 | end_vae = torch.cuda.Event(enable_timing=True) 211 | 212 | # VAE decode -> upscale -> encode 213 | start_vae.record() 214 | decoded_img = vae.decode(latents)[0] 215 | img_upscaled = torch.nn.functional.interpolate( 216 | decoded_img, scale_factor=args.scale, mode="bilinear" 217 | ) 218 | vae_encoded = vae.encode(img_upscaled).latent_dist.sample() 219 | end_vae.record() 220 | torch.cuda.synchronize() 221 | vae_elapsed = start_vae.elapsed_time(end_vae) 222 | 223 | # Scale latent 224 | start_latent = torch.cuda.Event(enable_timing=True) 225 | end_latent = torch.cuda.Event(enable_timing=True) 226 | start_latent.record() 227 | resized_latent = torch.nn.functional.interpolate( 228 | latents, scale_factor=args.scale, mode="bilinear" 229 | ) 230 | end_latent.record() 231 | torch.cuda.synchronize() 232 | latent_elapsed = start_latent.elapsed_time(end_latent) 233 | 234 | elapsed_vae_list.append(vae_elapsed) 235 | elapsed_latent_list.append(latent_elapsed) 236 | 237 | # Decode latents and calculate LPIPS and MSE 238 | decoded_vae = vae.decode(vae_encoded)[0] 239 | decoded_latent = vae.decode(resized_latent)[0] 240 | 241 | mse_vae = torch.nn.functional.mse_loss(decoded_vae, images_upscaled) 242 | mse_latent = torch.nn.functional.mse_loss( 243 | decoded_latent, images_upscaled 244 | ) 245 | 246 | mse_vae_list.extend(mse_vae.cpu().numpy().flatten()) 247 | mse_latent_list.extend(mse_latent.cpu().numpy().flatten()) 248 | 249 | lpips_vae = lpips_fn(decoded_vae.float(), images_upscaled.float()) 250 | lpips_latent = lpips_fn( 251 | decoded_latent.float(), images_upscaled.float() 252 | ) 253 | 254 | lpips_vae_list.extend(lpips_vae.cpu().numpy().flatten()) 255 | lpips_latent_list.extend(lpips_latent.cpu().numpy().flatten()) 256 | 257 | psnr_vae = psnr(decoded_vae, images_upscaled) 258 | psnr_latent = psnr(decoded_latent, images_upscaled) 259 | 260 | psnr_vae_list.extend(psnr_vae.cpu().numpy().flatten()) 261 | psnr_latent_list.extend(psnr_latent.cpu().numpy().flatten()) 262 | 263 | ssim_vae = ssim( 264 | 0.5 * (decoded_vae + 1), 265 | 0.5 * (images_upscaled + 1), 266 | data_range=1, 267 | size_average=True, 268 | ) 269 | ssim_latent = ssim( 270 | 0.5 * (decoded_latent + 1), 271 | 0.5 * (images_upscaled + 1), 272 | data_range=1, 273 | size_average=True, 274 | ) 275 | 276 | ssim_vae_list.append(ssim_vae.cpu()) 277 | ssim_latent_list.append(ssim_latent.cpu()) 278 | finally: 279 | print("Batch size", args.batch_size) 280 | if not args.resizer_only: 281 | print("Elapsed VAE", np.mean(elapsed_vae_list), "ms") 282 | print("Elapsed latent upscale", np.mean(elapsed_latent_list), "ms") 283 | print("Elapsed resizer", np.mean(elapsed_resizer_list), "ms") 284 | 285 | if not args.resizer_only: 286 | print("MSE VAE", np.mean(mse_vae_list)) 287 | print("MSE latent upscale", np.mean(mse_latent_list)) 288 | print("MSE resizer", np.mean(mse_resizer_list)) 289 | 290 | if not args.resizer_only: 291 | print("LPIPS VAE", np.mean(lpips_vae_list)) 292 | print("LPIPS latent upscale", np.mean(lpips_latent_list)) 293 | print("LPIPS resizer", np.mean(lpips_resizer_list)) 294 | 295 | if not args.resizer_only: 296 | print("PSNR VAE", np.mean(psnr_vae_list)) 297 | print("PSNR latent upscale", np.mean(psnr_latent_list)) 298 | print("PSNR resizer", np.mean(psnr_resizer_list)) 299 | 300 | if not args.resizer_only: 301 | print("SSIM VAE", np.mean(ssim_vae_list)) 302 | print("SSIM latent upscale", np.mean(ssim_latent_list)) 303 | print("SSIM resizer", np.mean(ssim_resizer_list)) 304 | -------------------------------------------------------------------------------- /examples/sdxl_kanagawa_512x512.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Ttl/ComfyUi_NNLatentUpscale/08105da31dbd7a54569661e135835e73bd8064b0/examples/sdxl_kanagawa_512x512.png -------------------------------------------------------------------------------- /examples/upscale.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Ttl/ComfyUi_NNLatentUpscale/08105da31dbd7a54569661e135835e73bd8064b0/examples/upscale.jpg -------------------------------------------------------------------------------- /latent_resizer.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | from einops import rearrange 6 | 7 | 8 | def normalization(channels): 9 | return nn.GroupNorm(32, channels) 10 | 11 | 12 | def zero_module(module): 13 | for p in module.parameters(): 14 | p.detach().zero_() 15 | return module 16 | 17 | 18 | class AttnBlock(nn.Module): 19 | def __init__(self, in_channels): 20 | super().__init__() 21 | self.in_channels = in_channels 22 | 23 | self.norm = normalization(in_channels) 24 | self.q = nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) 25 | self.k = nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) 26 | self.v = nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) 27 | self.proj_out = nn.Conv2d( 28 | in_channels, in_channels, kernel_size=1, stride=1, padding=0 29 | ) 30 | 31 | def attention(self, h_: torch.Tensor) -> torch.Tensor: 32 | h_ = self.norm(h_) 33 | q = self.q(h_) 34 | k = self.k(h_) 35 | v = self.v(h_) 36 | 37 | b, c, h, w = q.shape 38 | q, k, v = map( 39 | lambda x: rearrange(x, "b c h w -> b 1 (h w) c").contiguous(), (q, k, v) 40 | ) 41 | h_ = nn.functional.scaled_dot_product_attention( 42 | q, k, v 43 | ) # scale is dim ** -0.5 per default 44 | 45 | return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b) 46 | 47 | def forward(self, x, **kwargs): 48 | h_ = x 49 | h_ = self.attention(h_) 50 | h_ = self.proj_out(h_) 51 | return x + h_ 52 | 53 | 54 | def make_attn(in_channels, attn_kwargs=None): 55 | return AttnBlock(in_channels) 56 | 57 | 58 | class ResBlockEmb(nn.Module): 59 | def __init__( 60 | self, 61 | channels, 62 | emb_channels, 63 | dropout=0, 64 | out_channels=None, 65 | use_conv=False, 66 | use_scale_shift_norm=False, 67 | kernel_size=3, 68 | exchange_temb_dims=False, 69 | skip_t_emb=False, 70 | ): 71 | super().__init__() 72 | self.channels = channels 73 | self.emb_channels = emb_channels 74 | self.dropout = dropout 75 | self.out_channels = out_channels or channels 76 | self.use_conv = use_conv 77 | self.use_scale_shift_norm = use_scale_shift_norm 78 | self.exchange_temb_dims = exchange_temb_dims 79 | 80 | padding = kernel_size // 2 81 | 82 | self.in_layers = nn.Sequential( 83 | normalization(channels), 84 | nn.SiLU(), 85 | nn.Conv2d(channels, self.out_channels, kernel_size, padding=padding), 86 | ) 87 | 88 | self.skip_t_emb = skip_t_emb 89 | self.emb_out_channels = ( 90 | 2 * self.out_channels if use_scale_shift_norm else self.out_channels 91 | ) 92 | if self.skip_t_emb: 93 | print(f"Skipping timestep embedding in {self.__class__.__name__}") 94 | assert not self.use_scale_shift_norm 95 | self.emb_layers = None 96 | self.exchange_temb_dims = False 97 | else: 98 | self.emb_layers = nn.Sequential( 99 | nn.SiLU(), 100 | nn.Linear( 101 | emb_channels, 102 | self.emb_out_channels, 103 | ), 104 | ) 105 | 106 | self.out_layers = nn.Sequential( 107 | normalization(self.out_channels), 108 | nn.SiLU(), 109 | nn.Dropout(p=dropout), 110 | zero_module( 111 | nn.Conv2d( 112 | self.out_channels, 113 | self.out_channels, 114 | kernel_size, 115 | padding=padding, 116 | ) 117 | ), 118 | ) 119 | 120 | if self.out_channels == channels: 121 | self.skip_connection = nn.Identity() 122 | elif use_conv: 123 | self.skip_connection = nn.Conv2d( 124 | channels, self.out_channels, kernel_size, padding=padding 125 | ) 126 | else: 127 | self.skip_connection = nn.Conv2d(channels, self.out_channels, 1) 128 | 129 | def forward(self, x, emb): 130 | h = self.in_layers(x) 131 | 132 | if self.skip_t_emb: 133 | emb_out = torch.zeros_like(h) 134 | else: 135 | emb_out = self.emb_layers(emb).type(h.dtype) 136 | while len(emb_out.shape) < len(h.shape): 137 | emb_out = emb_out[..., None] 138 | if self.use_scale_shift_norm: 139 | out_norm, out_rest = self.out_layers[0], self.out_layers[1:] 140 | scale, shift = torch.chunk(emb_out, 2, dim=1) 141 | h = out_norm(h) * (1 + scale) + shift 142 | h = out_rest(h) 143 | else: 144 | if self.exchange_temb_dims: 145 | emb_out = rearrange(emb_out, "b t c ... -> b c t ...") 146 | h = h + emb_out 147 | h = self.out_layers(h) 148 | return self.skip_connection(x) + h 149 | 150 | 151 | class LatentResizer(nn.Module): 152 | def __init__(self, in_blocks=10, out_blocks=10, channels=128, dropout=0, attn=True): 153 | super().__init__() 154 | self.conv_in = nn.Conv2d(4, channels, 3, padding=1) 155 | 156 | self.channels = channels 157 | embed_dim = 32 158 | self.embed = nn.Sequential( 159 | nn.Linear(1, embed_dim), 160 | nn.SiLU(), 161 | nn.Linear(embed_dim, embed_dim), 162 | ) 163 | 164 | self.in_blocks = nn.ModuleList([]) 165 | for b in range(in_blocks): 166 | if (b == 1 or b == in_blocks - 1) and attn: 167 | self.in_blocks.append(make_attn(channels)) 168 | self.in_blocks.append(ResBlockEmb(channels, embed_dim, dropout)) 169 | 170 | self.out_blocks = nn.ModuleList([]) 171 | for b in range(out_blocks): 172 | if (b == 1 or b == out_blocks - 1) and attn: 173 | self.out_blocks.append(make_attn(channels)) 174 | self.out_blocks.append(ResBlockEmb(channels, embed_dim, dropout)) 175 | 176 | self.norm_out = normalization(channels) 177 | self.conv_out = nn.Conv2d(channels, 4, 3, padding=1) 178 | 179 | @classmethod 180 | def load_model(cls, filename, device="cpu", dtype=torch.float32, dropout=0): 181 | if not 'weights_only' in torch.load.__code__.co_varnames: 182 | weights = torch.load(filename, map_location=torch.device("cpu")) 183 | else: 184 | weights = torch.load(filename, map_location=torch.device("cpu"), weights_only=True) 185 | in_blocks = 0 186 | out_blocks = 0 187 | in_tfs = 0 188 | out_tfs = 0 189 | channels = weights["conv_in.bias"].shape[0] 190 | for k in weights.keys(): 191 | k = k.split(".") 192 | if k[0] == "in_blocks": 193 | in_blocks = max(in_blocks, int(k[1])) 194 | if k[2] == "q" and k[3] == "weight": 195 | in_tfs += 1 196 | if k[0] == "out_blocks": 197 | out_blocks = max(out_blocks, int(k[1])) 198 | if k[2] == "q" and k[3] == "weight": 199 | out_tfs += 1 200 | in_blocks = in_blocks + 1 - in_tfs 201 | out_blocks = out_blocks + 1 - out_tfs 202 | resizer = cls( 203 | in_blocks=in_blocks, 204 | out_blocks=out_blocks, 205 | channels=channels, 206 | dropout=dropout, 207 | attn=(out_tfs != 0), 208 | ) 209 | resizer.load_state_dict(weights) 210 | resizer.eval() 211 | resizer.to(device, dtype=dtype) 212 | return resizer 213 | 214 | def forward(self, x, scale=None, size=None): 215 | if scale is None and size is None: 216 | raise ValueError("Either scale or size needs to be not None") 217 | if scale is not None and size is not None: 218 | raise ValueError("Both scale or size can't be not None") 219 | if scale is not None: 220 | size = (x.shape[-2] * scale, x.shape[-1] * scale) 221 | size = tuple([int(round(i)) for i in size]) 222 | else: 223 | scale = size[-1] / x.shape[-1] 224 | 225 | # Output is the same size as input 226 | if size == x.shape[-2:]: 227 | return x 228 | 229 | scale = torch.tensor([scale - 1], dtype=x.dtype).to(x.device).unsqueeze(0) 230 | emb = self.embed(scale) 231 | 232 | x = self.conv_in(x) 233 | 234 | for b in self.in_blocks: 235 | if isinstance(b, ResBlockEmb): 236 | x = b(x, emb) 237 | else: 238 | x = b(x) 239 | x = F.interpolate(x, size=size, mode="bilinear") 240 | for b in self.out_blocks: 241 | if isinstance(b, ResBlockEmb): 242 | x = b(x, emb) 243 | else: 244 | x = b(x) 245 | 246 | x = self.norm_out(x) 247 | x = F.silu(x) 248 | x = self.conv_out(x) 249 | return x 250 | -------------------------------------------------------------------------------- /latent_resizer_train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | from torch.utils.data import DataLoader 6 | from torch.utils.data import Dataset 7 | from torch.utils.tensorboard import SummaryWriter 8 | from diffusers import AutoencoderKL 9 | import argparse 10 | import os 11 | import random 12 | import webdataset as wds 13 | import io 14 | from tqdm.auto import tqdm 15 | from latent_resizer import LatentResizer 16 | import lpips 17 | from collections import defaultdict 18 | from PIL import Image 19 | from torchvision import transforms 20 | 21 | 22 | def init_dataset(dataset_path, size=512): 23 | shards = [] 24 | if type(dataset_path) not in (list, tuple): 25 | dataset_path = [dataset_path] 26 | for path in dataset_path: 27 | for filename in os.listdir(path): 28 | full_path = os.path.join(path, filename) 29 | if full_path.endswith(".tar"): 30 | shards.append(full_path) 31 | print(f"{len(shards)} shards") 32 | 33 | def preprocess(sample): 34 | k = [k for k in sample.keys() if k in ["jpg", "png"]] 35 | if len(k) == 0: 36 | raise ValueError("Dataset images should be in jpg or png format") 37 | k = k[0] 38 | img = sample[k] 39 | 40 | img = Image.open(io.BytesIO(img)) 41 | if not img.mode == "RGB": 42 | img = img.convert("RGB") 43 | 44 | pil_image = img 45 | 46 | image_transforms = transforms.Compose( 47 | [ 48 | transforms.RandomCrop(size, pad_if_needed=True, padding_mode="reflect"), 49 | transforms.ToTensor(), 50 | transforms.Normalize([0.5], [0.5]), 51 | ] 52 | ) 53 | 54 | img = image_transforms(img) 55 | examples = {} 56 | examples["img"] = img 57 | return examples 58 | 59 | dataset = ( 60 | wds.WebDataset(shards, handler=wds.warn_and_continue, shardshuffle=True) 61 | .repeat() 62 | .shuffle(256) 63 | .map(preprocess) 64 | ) 65 | return dataset 66 | 67 | 68 | def collate_fn(examples): 69 | imgs = [example["img"] for example in examples] 70 | imgs = torch.stack(imgs) 71 | 72 | scale = random.uniform(1.0, 2.1) 73 | size = [int(round(imgs.shape[-2] * scale)), int(round(imgs.shape[-1] * scale))] 74 | size[0] -= size[0] % 8 75 | size[1] -= size[1] % 8 76 | imgs_scaled = F.interpolate(imgs, size=size, mode="bilinear") 77 | 78 | if scale < 1: 79 | batch = { 80 | "img_input": imgs_scaled, 81 | "img_target": imgs, 82 | } 83 | else: 84 | batch = { 85 | "img_input": imgs, 86 | "img_target": imgs_scaled, 87 | } 88 | 89 | return batch 90 | 91 | 92 | def calculate_loss( 93 | model, 94 | batch, 95 | vae, 96 | lpips, 97 | dtype=torch.float16, 98 | mse_weight=1, 99 | lpips_weight=0.1, 100 | mse_latent_weight=0.01, 101 | ): 102 | img_input = batch["img_input"].to(args.device, dtype=dtype) 103 | img_target = batch["img_target"].to(args.device, dtype=dtype) 104 | latent_input = ( 105 | vae.config.scaling_factor * vae.encode(img_input).latent_dist.sample() 106 | ) 107 | latent_target = ( 108 | vae.config.scaling_factor * vae.encode(img_target).latent_dist.sample() 109 | ) 110 | size = latent_target.shape[-2:] 111 | with torch.autocast(args.device, dtype=dtype, enabled=dtype != torch.float32): 112 | resized = model(latent_input, size=size) 113 | mse_latent = F.mse_loss(resized, latent_target) 114 | logs = {"mse_latent": mse_latent.detach().cpu().item()} 115 | decoded = vae.decode(resized / vae.config.scaling_factor)[0] 116 | mse = F.mse_loss(decoded, img_target) 117 | logs["mse"] = mse 118 | loss = mse_weight * mse + mse_latent_weight * mse_latent 119 | if lpips_weight > 0: 120 | ploss = lpips(decoded, img_target).mean() 121 | logs["lpips"] = ploss.detach().cpu().item() 122 | loss = loss + lpips_weight * ploss 123 | logs["loss"] = loss.detach().cpu().item() 124 | return loss, logs 125 | 126 | 127 | if __name__ == "__main__": 128 | parser = argparse.ArgumentParser(description="Latent interpolate trainer") 129 | parser.add_argument( 130 | "--train_path", 131 | required=True, 132 | action="append", 133 | help="Training data path for VAE latents. Webdataset format.", 134 | ) 135 | parser.add_argument( 136 | "--test_path", 137 | default=None, 138 | required=False, 139 | action="append", 140 | help="Test data path for VAE latents. Webdataset format.", 141 | ) 142 | parser.add_argument( 143 | "--vae_path", 144 | type=str, 145 | required=True, 146 | help="Path to VAE", 147 | ) 148 | parser.add_argument( 149 | "--test_steps", 150 | type=int, 151 | default=1000, 152 | required=False, 153 | help="Test interval", 154 | ) 155 | parser.add_argument( 156 | "--test_batches", 157 | type=int, 158 | default=10, 159 | required=False, 160 | help="Number of test batches", 161 | ) 162 | parser.add_argument( 163 | "--output_filename", 164 | type=str, 165 | default="sdxl_resizer.pt", 166 | required=False, 167 | help="Output filename", 168 | ) 169 | parser.add_argument( 170 | "--steps", 171 | type=int, 172 | default=100000, 173 | help="Number of steps to train", 174 | ) 175 | parser.add_argument( 176 | "--save_steps", 177 | type=int, 178 | default=5000, 179 | help="Save model every this step", 180 | ) 181 | parser.add_argument( 182 | "--batch_size", 183 | type=int, 184 | default=4, 185 | help="Batch size", 186 | ) 187 | parser.add_argument( 188 | "--num_workers", 189 | type=int, 190 | default=4, 191 | help="CPU workers", 192 | ) 193 | parser.add_argument( 194 | "--lr", 195 | type=float, 196 | default=2e-4, 197 | help="Learning rate", 198 | ) 199 | parser.add_argument( 200 | "--dropout", 201 | type=float, 202 | default=0.0, 203 | help="Droput rate", 204 | ) 205 | parser.add_argument( 206 | "--grad_clip", 207 | type=float, 208 | default=5.0, 209 | help="Gradient clipping", 210 | ) 211 | parser.add_argument( 212 | "--device", 213 | type=str, 214 | default="cuda", 215 | help="Device to use", 216 | ) 217 | parser.add_argument( 218 | "--resolution", 219 | type=int, 220 | default=256, 221 | help="Image resolution", 222 | ) 223 | parser.add_argument( 224 | "--init_weights", 225 | type=str, 226 | default=None, 227 | help="Resume training from weights file", 228 | ) 229 | parser.add_argument( 230 | "--fp16", 231 | action="store_true", 232 | help="Use fp16 precision", 233 | ) 234 | parser.add_argument( 235 | "--gradient_checkpointing", 236 | action="store_true", 237 | help="Enable gradient checkpointing for VAE", 238 | ) 239 | args = parser.parse_args() 240 | device = torch.device(args.device) 241 | 242 | vae_dtype = torch.float32 243 | if args.fp16: 244 | vae_dtype = torch.float16 245 | 246 | vae = AutoencoderKL.from_single_file(args.vae_path).to(device, dtype=vae_dtype) 247 | # Use this scale even with SD 1.5 248 | vae.config.scaling_factor = 0.13025 249 | 250 | vae.train() 251 | if args.gradient_checkpointing: 252 | vae.enable_gradient_checkpointing() 253 | lpips_fn = lpips.LPIPS(net="vgg").to(device=device, dtype=vae_dtype) 254 | 255 | if args.init_weights: 256 | model = LatentResizer.load_model( 257 | args.init_weights, 258 | device=args.device, 259 | dropout=args.dropout, 260 | dtype=torch.float32, 261 | ) 262 | else: 263 | model = LatentResizer(dropout=args.dropout).to(args.device) 264 | 265 | train_dataset = init_dataset(args.train_path, size=args.resolution) 266 | train_dataloader = DataLoader( 267 | train_dataset, 268 | batch_size=args.batch_size, 269 | collate_fn=collate_fn, 270 | num_workers=args.num_workers, 271 | ) 272 | if args.test_path: 273 | test_dataset = init_dataset(args.test_path, size=args.resolution) 274 | test_dataloader = DataLoader( 275 | test_dataset, 276 | batch_size=args.batch_size, 277 | collate_fn=collate_fn, 278 | num_workers=args.num_workers, 279 | ) 280 | 281 | optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr) 282 | scaler = torch.cuda.amp.GradScaler() 283 | scheduler1 = torch.optim.lr_scheduler.LinearLR( 284 | optimizer, start_factor=0.001, total_iters=200 285 | ) 286 | scheduler2 = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.steps) 287 | scheduler = torch.optim.lr_scheduler.SequentialLR( 288 | optimizer, schedulers=[scheduler1, scheduler2], milestones=[20] 289 | ) 290 | params = 0 291 | for p in model.parameters(): 292 | params += p.numel() 293 | print(params, "Parameters") 294 | writer = SummaryWriter(comment="resizer") 295 | model.train() 296 | epoch = 0 297 | step = 0 298 | progress_bar = tqdm(range(args.steps)) 299 | progress_bar.set_description("Steps") 300 | train_fn = lambda batch: calculate_loss(model, batch, vae, lpips_fn, vae_dtype) 301 | 302 | while step < args.steps: 303 | epoch += 1 304 | for batch in train_dataloader: 305 | if batch["img_input"].shape == batch["img_target"].shape: 306 | continue 307 | step += 1 308 | loss, logs = train_fn(batch) 309 | l = loss.detach().cpu().item() 310 | for k in logs.keys(): 311 | writer.add_scalar("{}/train".format(k), logs[k], step) 312 | progress_bar.set_postfix(loss=round(l, 2), lr=scheduler.get_last_lr()[0]) 313 | scaler.scale(loss).backward() 314 | if 0: 315 | total_norm = 0 316 | for p in model.parameters(): 317 | param_norm = p.grad.data.norm(2) 318 | total_norm += param_norm.item() ** 2 319 | total_norm = total_norm ** (1.0 / 2) 320 | print("norm", total_norm) 321 | if args.grad_clip > 0: 322 | nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip) 323 | scaler.step(optimizer) 324 | optimizer.zero_grad() 325 | scaler.update() 326 | progress_bar.update(1) 327 | scheduler.step() 328 | if step >= args.steps: 329 | break 330 | if (step % args.save_steps) == 0: 331 | base, ext = os.path.splitext(args.output_filename) 332 | save_filename = f"{base}-{step}{ext}" 333 | torch.save(model.state_dict(), save_filename) 334 | if args.test_path and (step % args.test_steps) == 0: 335 | test_batches = 0 336 | test_logs = defaultdict(float) 337 | test_loss = 0 338 | model.eval() 339 | for batch in test_dataloader: 340 | with torch.inference_mode(): 341 | _, logs = loss, logs = train_fn(batch) 342 | test_batches += 1 343 | for k in logs.keys(): 344 | test_logs[k] += logs[k] 345 | if test_batches >= args.test_batches: 346 | break 347 | model.train() 348 | for k in test_logs.keys(): 349 | writer.add_scalar( 350 | "{}/test".format(k), test_logs[k] / test_batches, step 351 | ) 352 | 353 | torch.save(model.state_dict(), args.output_filename) 354 | print("Model saved") 355 | -------------------------------------------------------------------------------- /nn_upscale.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from .latent_resizer import LatentResizer 3 | from comfy import model_management 4 | import os 5 | 6 | 7 | class NNLatentUpscale: 8 | """ 9 | Upscales SDXL latent using neural network 10 | """ 11 | 12 | def __init__(self): 13 | self.local_dir = os.path.dirname(os.path.realpath(__file__)) 14 | self.scale_factor = 0.13025 15 | self.dtype = torch.float32 16 | if model_management.should_use_fp16(): 17 | self.dtype = torch.float16 18 | self.weight_path = { 19 | "SDXL": os.path.join(self.local_dir, "sdxl_resizer.pt"), 20 | "SD 1.x": os.path.join(self.local_dir, "sd15_resizer.pt"), 21 | } 22 | self.version = "none" 23 | 24 | @classmethod 25 | def INPUT_TYPES(s): 26 | return { 27 | "required": { 28 | "latent": ("LATENT",), 29 | "version": (["SDXL", "SD 1.x"],), 30 | "upscale": ( 31 | "FLOAT", 32 | { 33 | "default": 1.5, 34 | "min": 1.0, 35 | "max": 2.0, 36 | "step": 0.01, 37 | "display": "number", 38 | }, 39 | ), 40 | }, 41 | } 42 | 43 | RETURN_TYPES = ("LATENT",) 44 | 45 | FUNCTION = "upscale" 46 | 47 | CATEGORY = "latent" 48 | 49 | def upscale(self, latent, version, upscale): 50 | device = model_management.get_torch_device() 51 | samples = latent["samples"].to(device=device, dtype=self.dtype) 52 | 53 | if version != self.version: 54 | self.model = LatentResizer.load_model( 55 | self.weight_path[version], device, self.dtype 56 | ) 57 | self.version = version 58 | 59 | self.model.to(device=device) 60 | latent_out = ( 61 | self.model(self.scale_factor * samples, scale=upscale) / self.scale_factor 62 | ) 63 | 64 | if self.dtype != torch.float32: 65 | latent_out = latent_out.to(dtype=torch.float32) 66 | 67 | latent_out = latent_out.to(device="cpu") 68 | 69 | self.model.to(device=model_management.vae_offload_device()) 70 | return ({"samples": latent_out},) 71 | -------------------------------------------------------------------------------- /sd15_resizer.pt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Ttl/ComfyUi_NNLatentUpscale/08105da31dbd7a54569661e135835e73bd8064b0/sd15_resizer.pt -------------------------------------------------------------------------------- /sdxl_resizer.pt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Ttl/ComfyUi_NNLatentUpscale/08105da31dbd7a54569661e135835e73bd8064b0/sdxl_resizer.pt --------------------------------------------------------------------------------