From 54edd29725a28fc47745869bdcef50d8c9b51002 Mon Sep 17 00:00:00 2001 From: lllyasviel Date: Fri, 9 Feb 2024 14:38:57 -0800 Subject: [PATCH] integrate latent modifier --- .../sd_forge_dynamic_thresholding/LICENSE.txt | 21 + .../sd_forge_latent_modifier/LICENSE | 674 ++++++++++ .../sd_forge_latent_modifier/README.md | 70 + .../sampler_mega_modifier.py | 1177 +++++++++++++++++ .../scripts/forge_latent_modifier.py | 104 ++ modules_forge/gradio_compile.py | 12 +- 6 files changed, 2057 insertions(+), 1 deletion(-) create mode 100644 extensions-builtin/sd_forge_dynamic_thresholding/LICENSE.txt create mode 100644 extensions-builtin/sd_forge_latent_modifier/LICENSE create mode 100644 extensions-builtin/sd_forge_latent_modifier/README.md create mode 100644 extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py create mode 100644 extensions-builtin/sd_forge_latent_modifier/scripts/forge_latent_modifier.py diff --git a/extensions-builtin/sd_forge_dynamic_thresholding/LICENSE.txt b/extensions-builtin/sd_forge_dynamic_thresholding/LICENSE.txt new file mode 100644 index 00000000..560a9bf3 --- /dev/null +++ b/extensions-builtin/sd_forge_dynamic_thresholding/LICENSE.txt @@ -0,0 +1,21 @@ +The MIT License (MIT) + +Copyright (c) 2023 Alex "mcmonkey" Goodwin + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/extensions-builtin/sd_forge_latent_modifier/LICENSE b/extensions-builtin/sd_forge_latent_modifier/LICENSE new file mode 100644 index 00000000..f288702d --- /dev/null +++ b/extensions-builtin/sd_forge_latent_modifier/LICENSE @@ -0,0 +1,674 @@ + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/extensions-builtin/sd_forge_latent_modifier/README.md b/extensions-builtin/sd_forge_latent_modifier/README.md new file mode 100644 index 00000000..d1b85da8 --- /dev/null +++ b/extensions-builtin/sd_forge_latent_modifier/README.md @@ -0,0 +1,70 @@ +This extension is compiled from https://github.com/Clybius +Original Licence GPL V3 + + +## Latent Diffusion Mega Modifier (sampler_mega_modifier.py) +### Adds multiple parameters to control the diffusion process towards a quality the user expects. +* Sharpness: utilizes code from Fooocus's sampling process to sharpen the noise in the middle of the diffusion process. +This can lead to more perceptual detail, especially at higher strengths. + +* Tonemap: Clamps conditioning noise (CFG) using a user-chosen method, which can allow for the use of higher CFG values. + +* Rescale: Scales the CFG by comparing the standard deviation to the existing latent to dynamically lower the CFG. + +* Extra Noise: Adds extra noise in the middle of the diffusion process to conditioning, and do the inverse operation on unconditioning, if chosen. + +* Contrast: Adjusts the contrast of the conditioning, can lead to more pop-style results. Essentially functions as a secondary CFG slider for stylization, without changing subject pose and location much, if at all. + +* Combat CFG Drift: As we increase CFG, the mean will slightly drift away from 0. This subtracts the mean or median of the latent. Can lead to potentially sharper and higher frequency results, but may result in discoloration. + +* Divisive Norm: Normalizes the latent using avg_pool2d, and can reduce noisy artifacts, due in part to features such as sharpness. + +* Spectral Modulation: Converts the latent to frequencies, and clamps higher frequencies while boosting lower ones, then converts it back to an image latent. This effectively can be treated as a solution to oversaturation or burning as a result of higher CFG values, while not touching values around the median. + +### Tonemapping Methods Explanation: +* Reinhard:

Uses the reinhard method of tonemapping (from comfyanonymous' ComfyUI Experiments) to clamp the CFG if the difference is too strong. + + Lower `tonemap_multiplier` clamps more noise, and a lower `tonemap_percentile` will increase the calculated standard deviation from the original noise. Play with it!

+* Arctan:

Clamps the values dynamically using a simple arctan curve. [Link to interactive Desmos visualization](https://www.desmos.com/calculator/e4nrcdpqbl). + + Recommended values for testing: tonemap_multiplier of 5, tonemap_percentile of 90.

+* Quantile:

Clamps the values using torch.quantile for obtaining the highest magnitudes, and clamping based on the result. + + + `Closer to 100 percentile == stronger clamping`. Recommended values for testing: tonemap_multiplier of 1, tonemap_percentile of 99.

+* Gated:

Clamps the values using torch.quantile, only if above a specific floor value, which is set by `tonemapping_multiplier`. Clamps the noise prediction latent based on the percentile. + + + `Closer to 100 percentile == stronger clamping, lower tonemapping_multiplier == stronger clamping`. Recommended values for testing: tonemap_multiplier of 0.8-1, tonemap_percentile of 99.995.

+* CFG-Mimic:

Attempts to mimic a lower or higher CFG based on `tonemapping_multiplier`, and clamps it using `tonemapping_percentile` with torch.quantile. + + + `Closer to 100 percentile == stronger clamping, lower tonemapping_multiplier == stronger clamping`. Recommended values for testing: tonemap_multiplier of 0.33-1.0, tonemap_percentile of 100.

+* Spatial-Norm:

Clamps the values according to the noise prediction's absolute mean in the spectral domain. `tonemap_multiplier` adjusts the strength of the clamping. + + + `Lower tonemapping_multiplier == stronger clamping`. Recommended value for testing: tonemap_multiplier of 0.5-2.0.

+ +### Contrast Explanation: +

Scales the pixel values by the standard deviation, achieving a more contrasty look. In practice, this can effectively act as a secondary CFG slider for stylization. It doesn't modify subject poses much, if at all, which can be great for those looking to get more oomf out of their low-cfg setups. + +Using a negative value will apply the inverse of the operation to the latent.

+ +### Spectral Modification Explanation: +

We boost the low frequencies (low rate of change in the noise), and we lower the high frequencies (high rates of change in the noise). + +Change the low/high frequency range using `spectral_mod_percentile` (default of 5.0, which is the upper and lower 5th percentiles.) + +Increase/Decrease the strength of the adjustment by increasing `spectral_mod_multiplier` + +Beware of percentile values higher than 15 and multiplier values higher than 5, especially for hard clamping. Here be dragons, as large values may cause it to "noise-out", or become full of non-sensical noise, especially earlier in the diffusion process.

+ + +#### Current Pipeline: +>##### Add extra noise to conditioning -> Sharpen conditioning -> Convert to Noise Prediction -> Tonemap Noise Prediction -> Spectral Modification -> Modify contrast of noise prediction -> Rescale CFG -> Divisive Normalization -> Combat CFG Drift + +#### Why use this over `x` node? +Since the `set_model_sampler_cfg_function` hijack in ComfyUI can only utilize a single function, we bundle many latent modification methods into one large function for processing. This is simpler than taking an existing hijack and modifying it, which may be possible, but my (Clybius') lack of Python/PyTorch knowledge leads to this being the optimal method for simplicity. If you know how to do this, feel free to reach out through any means! + +#### Can you implement `x` function? +Depends. Is there existing code for such a function, with an open license for possible use in this repository? I could likely attempt adding it! Feel free to start an issue or to reach out for ideas you'd want implemented. diff --git a/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py b/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py new file mode 100644 index 00000000..f8b494e2 --- /dev/null +++ b/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py @@ -0,0 +1,1177 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import numpy as np +import random + +# Set manual seeds for noise +# rand(n)_like. but with generator support +def gen_like(f, input, generator=None): + return f(input.size(), generator=generator).to(input) + +''' + The following snippet is utilized from https://github.com/Jamy-L/Pytorch-Contrast-Adaptive-Sharpening/ +''' +def min_(tensor_list): + # return the element-wise min of the tensor list. + x = torch.stack(tensor_list) + mn = x.min(axis=0)[0] + return mn#torch.clamp(mn, min=-1) + +def max_(tensor_list): + # return the element-wise max of the tensor list. + x = torch.stack(tensor_list) + mx = x.max(axis=0)[0] + return mx#torch.clamp(mx, max=1) +def contrast_adaptive_sharpening(image, amount): + img = F.pad(image, pad=(1, 1, 1, 1)) + absmean = torch.abs(image.mean()) + + a = img[..., :-2, :-2] + b = img[..., :-2, 1:-1] + c = img[..., :-2, 2:] + d = img[..., 1:-1, :-2] + e = img[..., 1:-1, 1:-1] + f = img[..., 1:-1, 2:] + g = img[..., 2:, :-2] + h = img[..., 2:, 1:-1] + i = img[..., 2:, 2:] + + # Computing contrast + cross = (b, d, e, f, h) + mn = min_(cross) + mx = max_(cross) + + diag = (a, c, g, i) + mn2 = min_(diag) + mx2 = max_(diag) + mx = mx + mx2 + mn = mn + mn2 + + # Computing local weight + inv_mx = torch.reciprocal(mx) + amp = inv_mx * torch.minimum(mn, (2 - mx)) + + # scaling + amp = torch.copysign(torch.sqrt(torch.abs(amp)), amp) + w = - amp * (amount * (1/5 - 1/8) + 1/8) + div = torch.reciprocal(1 + 4*w).clamp(-10, 10) + + output = ((b + d + f + h)*w + e) * div + output = torch.nan_to_num(output) + + return (output.to(image.device)) + +''' + The following gaussian functions were utilized from the Fooocus UI, many thanks to github.com/Illyasviel ! +''' +def gaussian_kernel(kernel_size, sigma): + kernel = np.fromfunction( + lambda x, y: (1 / (2 * np.pi * sigma ** 2)) * + np.exp(-((x - (kernel_size - 1) / 2) ** 2 + (y - (kernel_size - 1) / 2) ** 2) / (2 * sigma ** 2)), + (kernel_size, kernel_size) + ) + return kernel / np.sum(kernel) + + +class GaussianBlur(nn.Module): + def __init__(self, channels, kernel_size, sigma): + super(GaussianBlur, self).__init__() + self.channels = channels + self.kernel_size = kernel_size + self.sigma = sigma + self.padding = kernel_size // 2 # Ensure output size matches input size + self.register_buffer('kernel', torch.tensor(gaussian_kernel(kernel_size, sigma), dtype=torch.float32)) + self.kernel = self.kernel.view(1, 1, kernel_size, kernel_size) + self.kernel = self.kernel.expand(self.channels, -1, -1, -1) # Repeat the kernel for each input channel + + def forward(self, x): + x = F.conv2d(x, self.kernel.to(x), padding=self.padding, groups=self.channels) + return x + +gaussian_filter_2d = GaussianBlur(4, 7, 0.8) + +''' + As of August 18th (on Fooocus' GitHub), the gaussian functions were replaced by an anisotropic function for better stability. +''' +Tensor = torch.Tensor +Device = torch.DeviceObjType +Dtype = torch.Type +pad = torch.nn.functional.pad + + +def _compute_zero_padding(kernel_size: tuple[int, int] | int) -> tuple[int, int]: + ky, kx = _unpack_2d_ks(kernel_size) + return (ky - 1) // 2, (kx - 1) // 2 + + +def _unpack_2d_ks(kernel_size: tuple[int, int] | int) -> tuple[int, int]: + if isinstance(kernel_size, int): + ky = kx = kernel_size + else: + assert len(kernel_size) == 2, '2D Kernel size should have a length of 2.' + ky, kx = kernel_size + + ky = int(ky) + kx = int(kx) + return ky, kx + + +def gaussian( + window_size: int, sigma: Tensor | float, *, device: Device | None = None, dtype: Dtype | None = None +) -> Tensor: + + batch_size = sigma.shape[0] + + x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1) + + if window_size % 2 == 0: + x = x + 0.5 + + gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0))) + + return gauss / gauss.sum(-1, keepdim=True) + + +def get_gaussian_kernel1d( + kernel_size: int, + sigma: float | Tensor, + force_even: bool = False, + *, + device: Device | None = None, + dtype: Dtype | None = None, +) -> Tensor: + + return gaussian(kernel_size, sigma, device=device, dtype=dtype) + + +def get_gaussian_kernel2d( + kernel_size: tuple[int, int] | int, + sigma: tuple[float, float] | Tensor, + force_even: bool = False, + *, + device: Device | None = None, + dtype: Dtype | None = None, +) -> Tensor: + + sigma = torch.Tensor([[sigma, sigma]]).to(device=device, dtype=dtype) + + ksize_y, ksize_x = _unpack_2d_ks(kernel_size) + sigma_y, sigma_x = sigma[:, 0, None], sigma[:, 1, None] + + kernel_y = get_gaussian_kernel1d(ksize_y, sigma_y, force_even, device=device, dtype=dtype)[..., None] + kernel_x = get_gaussian_kernel1d(ksize_x, sigma_x, force_even, device=device, dtype=dtype)[..., None] + + return kernel_y * kernel_x.view(-1, 1, ksize_x) + + +def _bilateral_blur( + input: Tensor, + guidance: Tensor | None, + kernel_size: tuple[int, int] | int, + sigma_color: float | Tensor, + sigma_space: tuple[float, float] | Tensor, + border_type: str = 'reflect', + color_distance_type: str = 'l1', +) -> Tensor: + + if isinstance(sigma_color, Tensor): + sigma_color = sigma_color.to(device=input.device, dtype=input.dtype).view(-1, 1, 1, 1, 1) + + ky, kx = _unpack_2d_ks(kernel_size) + pad_y, pad_x = _compute_zero_padding(kernel_size) + + padded_input = pad(input, (pad_x, pad_x, pad_y, pad_y), mode=border_type) + unfolded_input = padded_input.unfold(2, ky, 1).unfold(3, kx, 1).flatten(-2) # (B, C, H, W, Ky x Kx) + + if guidance is None: + guidance = input + unfolded_guidance = unfolded_input + else: + padded_guidance = pad(guidance, (pad_x, pad_x, pad_y, pad_y), mode=border_type) + unfolded_guidance = padded_guidance.unfold(2, ky, 1).unfold(3, kx, 1).flatten(-2) # (B, C, H, W, Ky x Kx) + + diff = unfolded_guidance - guidance.unsqueeze(-1) + if color_distance_type == "l1": + color_distance_sq = diff.abs().sum(1, keepdim=True).square() + elif color_distance_type == "l2": + color_distance_sq = diff.square().sum(1, keepdim=True) + else: + raise ValueError("color_distance_type only acceps l1 or l2") + color_kernel = (-0.5 / sigma_color**2 * color_distance_sq).exp() # (B, 1, H, W, Ky x Kx) + + space_kernel = get_gaussian_kernel2d(kernel_size, sigma_space, device=input.device, dtype=input.dtype) + space_kernel = space_kernel.view(-1, 1, 1, 1, kx * ky) + + kernel = space_kernel * color_kernel + out = (unfolded_input * kernel).sum(-1) / kernel.sum(-1) + return out + + +def bilateral_blur( + input: Tensor, + kernel_size: tuple[int, int] | int = (13, 13), + sigma_color: float | Tensor = 3.0, + sigma_space: tuple[float, float] | Tensor = 3.0, + border_type: str = 'reflect', + color_distance_type: str = 'l1', +) -> Tensor: + return _bilateral_blur(input, None, kernel_size, sigma_color, sigma_space, border_type, color_distance_type) + + +def joint_bilateral_blur( + input: Tensor, + guidance: Tensor, + kernel_size: tuple[int, int] | int, + sigma_color: float | Tensor, + sigma_space: tuple[float, float] | Tensor, + border_type: str = 'reflect', + color_distance_type: str = 'l1', +) -> Tensor: + return _bilateral_blur(input, guidance, kernel_size, sigma_color, sigma_space, border_type, color_distance_type) + + +class _BilateralBlur(torch.nn.Module): + def __init__( + self, + kernel_size: tuple[int, int] | int, + sigma_color: float | Tensor, + sigma_space: tuple[float, float] | Tensor, + border_type: str = 'reflect', + color_distance_type: str = "l1", + ) -> None: + super().__init__() + self.kernel_size = kernel_size + self.sigma_color = sigma_color + self.sigma_space = sigma_space + self.border_type = border_type + self.color_distance_type = color_distance_type + + def __repr__(self) -> str: + return ( + f"{self.__class__.__name__}" + f"(kernel_size={self.kernel_size}, " + f"sigma_color={self.sigma_color}, " + f"sigma_space={self.sigma_space}, " + f"border_type={self.border_type}, " + f"color_distance_type={self.color_distance_type})" + ) + + +class BilateralBlur(_BilateralBlur): + def forward(self, input: Tensor) -> Tensor: + return bilateral_blur( + input, self.kernel_size, self.sigma_color, self.sigma_space, self.border_type, self.color_distance_type + ) + + +class JointBilateralBlur(_BilateralBlur): + def forward(self, input: Tensor, guidance: Tensor) -> Tensor: + return joint_bilateral_blur( + input, + guidance, + self.kernel_size, + self.sigma_color, + self.sigma_space, + self.border_type, + self.color_distance_type, + ) + + +# Below is perlin noise from https://github.com/tasptz/pytorch-perlin-noise/blob/main/perlin_noise/perlin_noise.py +from torch import Generator, Tensor, lerp +from torch.nn.functional import unfold +from typing import Callable, Tuple +from math import pi + +def get_positions(block_shape: Tuple[int, int]) -> Tensor: + """ + Generate position tensor. + + Arguments: + block_shape -- (height, width) of position tensor + + Returns: + position vector shaped (1, height, width, 1, 1, 2) + """ + bh, bw = block_shape + positions = torch.stack( + torch.meshgrid( + [(torch.arange(b) + 0.5) / b for b in (bw, bh)], + indexing="xy", + ), + -1, + ).view(1, bh, bw, 1, 1, 2) + return positions + + +def unfold_grid(vectors: Tensor) -> Tensor: + """ + Unfold vector grid to batched vectors. + + Arguments: + vectors -- grid vectors + + Returns: + batched grid vectors + """ + batch_size, _, gpy, gpx = vectors.shape + return ( + unfold(vectors, (2, 2)) + .view(batch_size, 2, 4, -1) + .permute(0, 2, 3, 1) + .view(batch_size, 4, gpy - 1, gpx - 1, 2) + ) + + +def smooth_step(t: Tensor) -> Tensor: + """ + Smooth step function [0, 1] -> [0, 1]. + + Arguments: + t -- input values (any shape) + + Returns: + output values (same shape as input values) + """ + return t * t * (3.0 - 2.0 * t) + + +def perlin_noise_tensor( + vectors: Tensor, positions: Tensor, step: Callable = None +) -> Tensor: + """ + Generate perlin noise from batched vectors and positions. + + Arguments: + vectors -- batched grid vectors shaped (batch_size, 4, grid_height, grid_width, 2) + positions -- batched grid positions shaped (batch_size or 1, block_height, block_width, grid_height or 1, grid_width or 1, 2) + + Keyword Arguments: + step -- smooth step function [0, 1] -> [0, 1] (default: `smooth_step`) + + Raises: + Exception: if position and vector shapes do not match + + Returns: + (batch_size, block_height * grid_height, block_width * grid_width) + """ + if step is None: + step = smooth_step + + batch_size = vectors.shape[0] + # grid height, grid width + gh, gw = vectors.shape[2:4] + # block height, block width + bh, bw = positions.shape[1:3] + + for i in range(2): + if positions.shape[i + 3] not in (1, vectors.shape[i + 2]): + raise Exception( + f"Blocks shapes do not match: vectors ({vectors.shape[1]}, {vectors.shape[2]}), positions {gh}, {gw})" + ) + + if positions.shape[0] not in (1, batch_size): + raise Exception( + f"Batch sizes do not match: vectors ({vectors.shape[0]}), positions ({positions.shape[0]})" + ) + + vectors = vectors.view(batch_size, 4, 1, gh * gw, 2) + positions = positions.view(positions.shape[0], bh * bw, -1, 2) + + step_x = step(positions[..., 0]) + step_y = step(positions[..., 1]) + + row0 = lerp( + (vectors[:, 0] * positions).sum(dim=-1), + (vectors[:, 1] * (positions - positions.new_tensor((1, 0)))).sum(dim=-1), + step_x, + ) + row1 = lerp( + (vectors[:, 2] * (positions - positions.new_tensor((0, 1)))).sum(dim=-1), + (vectors[:, 3] * (positions - positions.new_tensor((1, 1)))).sum(dim=-1), + step_x, + ) + noise = lerp(row0, row1, step_y) + return ( + noise.view( + batch_size, + bh, + bw, + gh, + gw, + ) + .permute(0, 3, 1, 4, 2) + .reshape(batch_size, gh * bh, gw * bw) + ) + + +def perlin_noise( + grid_shape: Tuple[int, int], + out_shape: Tuple[int, int], + batch_size: int = 1, + generator: Generator = None, + *args, + **kwargs, +) -> Tensor: + """ + Generate perlin noise with given shape. `*args` and `**kwargs` are forwarded to `Tensor` creation. + + Arguments: + grid_shape -- Shape of grid (height, width). + out_shape -- Shape of output noise image (height, width). + + Keyword Arguments: + batch_size -- (default: {1}) + generator -- random generator used for grid vectors (default: {None}) + + Raises: + Exception: if grid and out shapes do not match + + Returns: + Noise image shaped (batch_size, height, width) + """ + # grid height and width + gh, gw = grid_shape + # output height and width + oh, ow = out_shape + # block height and width + bh, bw = oh // gh, ow // gw + + if oh != bh * gh: + raise Exception(f"Output height {oh} must be divisible by grid height {gh}") + if ow != bw * gw != 0: + raise Exception(f"Output width {ow} must be divisible by grid width {gw}") + + angle = torch.empty( + [batch_size] + [s + 1 for s in grid_shape], *args, **kwargs + ).uniform_(to=2.0 * pi, generator=generator) + # random vectors on grid points + vectors = unfold_grid(torch.stack((torch.cos(angle), torch.sin(angle)), dim=1)) + # positions inside grid cells [0, 1) + positions = get_positions((bh, bw)).to(vectors) + return perlin_noise_tensor(vectors, positions).squeeze(0) + +def generate_1f_noise(tensor, alpha, k, generator=None): + """Generate 1/f noise for a given tensor. + + Args: + tensor: The tensor to add noise to. + alpha: The parameter that determines the slope of the spectrum. + k: A constant. + + Returns: + A tensor with the same shape as `tensor` containing 1/f noise. + """ + fft = torch.fft.fft2(tensor) + freq = torch.arange(1, len(fft) + 1, dtype=torch.float) + spectral_density = k / freq**alpha + noise = torch.randn(tensor.shape, generator=generator) * spectral_density + return noise + +def green_noise(width, height, generator=None): + noise = torch.randn(width, height, generator=generator) + scale = 1.0 / (width * height) + fy = torch.fft.fftfreq(width)[:, None] ** 2 + fx = torch.fft.fftfreq(height) ** 2 + f = fy + fx + power = torch.sqrt(f) + power[0, 0] = 1 + noise = torch.fft.ifft2(torch.fft.fft2(noise) / torch.sqrt(power)) + noise *= scale / noise.std() + return torch.real(noise) + +# Algorithm from https://github.com/v0xie/sd-webui-cads/ +def add_cads_noise(y, timestep, cads_schedule_start, cads_schedule_end, cads_noise_scale, cads_rescale_factor, cads_rescale=False): + timestep_as_float = (timestep / 999.0)[:, None, None, None].clone()[0].item() + gamma = 0.0 + if timestep_as_float < cads_schedule_start: + gamma = 1.0 + elif timestep_as_float > cads_schedule_end: + gamma = 0.0 + else: + gamma = (cads_schedule_end - timestep_as_float) / (cads_schedule_end - cads_schedule_start) + + y_mean, y_std = torch.mean(y), torch.std(y) + y = np.sqrt(gamma) * y + cads_noise_scale * np.sqrt(1 - gamma) * torch.randn_like(y) + + if cads_rescale: + y_scaled = (y - torch.mean(y)) / torch.std(y) * y_std + y_mean + if not torch.isnan(y_scaled).any(): + y = cads_rescale_factor * y_scaled + (1 - cads_rescale_factor) * y + else: + print("Encountered NaN in cads rescaling. Skipping rescaling.") + return y + +# Algorithm from https://github.com/v0xie/sd-webui-cads/ +def add_cads_custom_noise(y, noise, timestep, cads_schedule_start, cads_schedule_end, cads_noise_scale, cads_rescale_factor, cads_rescale=False): + timestep_as_float = (timestep / 999.0)[:, None, None, None].clone()[0].item() + gamma = 0.0 + if timestep_as_float < cads_schedule_start: + gamma = 1.0 + elif timestep_as_float > cads_schedule_end: + gamma = 0.0 + else: + gamma = (cads_schedule_end - timestep_as_float) / (cads_schedule_end - cads_schedule_start) + + y_mean, y_std = torch.mean(y), torch.std(y) + y = np.sqrt(gamma) * y + cads_noise_scale * np.sqrt(1 - gamma) * noise#.sub_(noise.mean()).div_(noise.std()) + + if cads_rescale: + y_scaled = (y - torch.mean(y)) / torch.std(y) * y_std + y_mean + if not torch.isnan(y_scaled).any(): + y = cads_rescale_factor * y_scaled + (1 - cads_rescale_factor) * y + else: + print("Encountered NaN in cads rescaling. Skipping rescaling.") + return y + +# Tonemapping functions + +def train_difference(a: Tensor, b: Tensor, c: Tensor) -> Tensor: + diff_AB = a.float() - b.float() + distance_A0 = torch.abs(b.float() - c.float()) + distance_A1 = torch.abs(b.float() - a.float()) + + sum_distances = distance_A0 + distance_A1 + + scale = torch.where( + sum_distances != 0, distance_A1 / sum_distances, torch.tensor(0.0).float() + ) + sign_scale = torch.sign(b.float() - c.float()) + scale = sign_scale * torch.abs(scale) + new_diff = scale * torch.abs(diff_AB) + return new_diff + +def gated_thresholding(percentile: float, floor: float, t: Tensor) -> Tensor: + """ + Args: + percentile: float between 0.0 and 1.0. for example 0.995 would subject only the top 0.5%ile to clamping. + t: [b, c, v] tensor in pixel or latent space (where v is the result of flattening w and h) + """ + a = t.abs() # Magnitudes + q = torch.quantile(a, percentile, dim=2) # Get clamp value via top % of magnitudes + q.clamp_(min=floor) + q = q.unsqueeze(2).expand(*t.shape) + t = t.clamp(-q, q) # Clamp latent with magnitude value + t = t / q + return t + +def dyn_thresh_gate(latent: Tensor, centered_magnitudes: Tensor, tonemap_percentile: float, floor: float, ceil: float): + if centered_magnitudes.lt(torch.tensor(ceil, device=centered_magnitudes.device)).all().item(): # If the magnitudes are less than the ceiling + return latent # Return the unmodified centered latent + else: + latent = gated_thresholding(tonemap_percentile, floor, latent) # If the magnitudes are higher than the ceiling + return latent # Gated-dynamic thresholding by Birchlabs + +def spatial_norm_thresholding(x0, value): + # b c h w + pow_x0 = torch.pow(torch.abs(x0), 2) + s = pow_x0.mean(1, keepdim=True).sqrt().clamp(min=value) + return x0 * (value / s) + +def spatial_norm_chw_thresholding(x0, value): + # b c h w + pow_x0 = torch.pow(torch.abs(x0), 2) + s = pow_x0.mean(dim=(1, 2, 3), keepdim=True).sqrt().clamp(min=value) + return x0 * (value / s) + +# Contrast function + +def contrast(x: Tensor): + # Calculate the mean and standard deviation of the pixel values + #mean = x.mean(dim=(1,2,3), keepdim=True) + stddev = x.std(dim=(1,2,3), keepdim=True) + # Scale the pixel values by the standard deviation + scaled_pixels = (x) / stddev + return scaled_pixels + +def contrast_with_mean(x: Tensor): + # Calculate the mean and standard deviation of the pixel values + #mean = x.mean(dim=(2,3), keepdim=True) + stddev = x.std(dim=(1,2,3), keepdim=True) + diff_mean = ((x / stddev) - x).mean(dim=(1,2,3), keepdim=True) + # Scale the pixel values by the standard deviation + scaled_pixels = x / stddev + return scaled_pixels - diff_mean + +def center_latent(tensor): #https://birchlabs.co.uk/machine-learning#combating-mean-drift-in-cfg + """Centers on 0 to combat CFG drift.""" + tensor = tensor - tensor.mean(dim=(-2, -1)).unsqueeze(-1).unsqueeze(-1).expand(tensor.shape) + return tensor + +def center_0channel(tensor): #https://birchlabs.co.uk/machine-learning#combating-mean-drift-in-cfg + """Centers on 0 to combat CFG drift.""" + std_dev_0 = tensor[:, [0]].std() + mean_0 = tensor[:, [0]].mean() + mean_12 = tensor[:, [1,2]].mean() + mean_3 = tensor[:, [3]].mean() + + #tensor[:, [0]] /= std_dev_0 + tensor[:, [0]] -= mean_0 + tensor[:, [0]] += torch.copysign(torch.pow(torch.abs(mean_0), 1.5), mean_0) + #tensor[:, [1, 2]] -= tensor[:, [1, 2]].mean() + tensor[:, [1, 2]] -= mean_12 * 0.5 + tensor[:, [3]] -= mean_3 + tensor[:, [3]] += torch.copysign(torch.pow(torch.abs(mean_3), 1.5), mean_3) + return tensor# - tensor.mean(dim=(2,3), keepdim=True) + +def channel_sharpen(tensor): + """Centers on 0 to combat CFG drift.""" + flattened = tensor.flatten(2) + flat_std = flattened.std(dim=(2)).unsqueeze(2).expand(flattened.shape) + flattened *= flat_std + flattened -= flattened.mean(dim=(2)).unsqueeze(2).expand(flattened.shape) + flattened /= flat_std + tensor = flattened.unflatten(2, tensor.shape[2:]) + return tensor + + +def center_012channel(tensor): #https://birchlabs.co.uk/machine-learning#combating-mean-drift-in-cfg + """Centers on 0 to combat CFG drift.""" + curr_tens = tensor[:, [0,1,2]] + tensor[:, [0,1,2]] -= curr_tens.mean() + return tensor + +def center_latent_perchannel(tensor): # Does nothing different than above + """Centers on 0 to combat CFG drift.""" + flattened = tensor.flatten(2) + flattened = flattened - flattened.mean(dim=(2)).unsqueeze(2).expand(flattened.shape) + tensor = flattened.unflatten(2, tensor.shape[2:]) + return tensor + +def center_latent_perchannel_with_magnitudes(tensor): # Does nothing different than above + """Centers on 0 to combat CFG drift.""" + flattened = tensor.flatten(2) + flattened_magnitude = (torch.linalg.vector_norm(flattened, dim=(2), keepdim=True) + 0.0000000001) + flattened /= flattened_magnitude + flattened = flattened - flattened.mean(dim=(2)).unsqueeze(2).expand(flattened.shape) + flattened *= flattened_magnitude + tensor = flattened.unflatten(2, tensor.shape[2:]) + return tensor + +def center_latent_perchannel_with_decorrelate(tensor): # Decorrelates data, slight change, test and play with it. + """Centers on 0 to combat CFG drift, preprocesses the latent with decorrelation""" + tensor = decorrelate_data(tensor) + flattened = tensor.flatten(2) + flattened_magnitude = (torch.linalg.vector_norm(flattened, dim=(2), keepdim=True) + 0.0000000001) + flattened /= flattened_magnitude + flattened = flattened - flattened.mean(dim=(2)).unsqueeze(2).expand(flattened.shape) + flattened *= flattened_magnitude + tensor = flattened.unflatten(2, tensor.shape[2:]) + return tensor + +def center_latent_median(tensor): + flattened = tensor.flatten(2) + median = flattened.median() + scaled_data = (flattened - median) + scaled_data = scaled_data.unflatten(2, tensor.shape[2:]) + return scaled_data + +def divisive_normalization(image_tensor, neighborhood_size, threshold=1e-6): + # Compute the local mean and local variance + local_mean = F.avg_pool2d(image_tensor, neighborhood_size, stride=1, padding=neighborhood_size // 2, count_include_pad=False) + local_mean_squared = local_mean**2 + + local_variance = F.avg_pool2d(image_tensor**2, neighborhood_size, stride=1, padding=neighborhood_size // 2, count_include_pad=False) - local_mean_squared + + # Add a small value to prevent division by zero + local_variance = local_variance + threshold + + # Apply divisive normalization + normalized_tensor = image_tensor / torch.sqrt(local_variance) + + return normalized_tensor + +def decorrelate_data(data): + """flattened = tensor.flatten(2).squeeze(0) # this code aint shit, yo + cov_matrix = torch.cov(flattened) + sqrt_inv_cov_matrix = torch.linalg.inv(torch.sqrt(cov_matrix)) + decorrelated_tensor = torch.dot(flattened, sqrt_inv_cov_matrix.T) + decorrelated_tensor = decorrelated_tensor.unflatten(2, tensor.shape[2:]).unsqueeze(0)""" + + # Reshape the 4D tensor to a 2D tensor for covariance calculation + num_samples, num_channels, height, width = data.size() + data_reshaped = data.view(num_samples, num_channels, -1) + data_reshaped = data_reshaped - torch.mean(data_reshaped, dim=2, keepdim=True) + + # Compute covariance matrix + cov_matrix = torch.matmul(data_reshaped, data_reshaped.transpose(1, 2)) / (height * width - 1) + + # Compute the inverse square root of the covariance matrix + u, s, v = torch.svd(cov_matrix) + sqrt_inv_cov_matrix = torch.matmul(u, torch.matmul(torch.diag_embed(1.0 / torch.sqrt(s)), v.transpose(1, 2))) + + # Reshape sqrt_inv_cov_matrix to match the dimensions of data_reshaped + sqrt_inv_cov_matrix = sqrt_inv_cov_matrix.unsqueeze(0).expand(num_samples, -1, -1, -1) + + # Decorrelate the data + decorrelated_data = torch.matmul(data_reshaped.transpose(1, 2), sqrt_inv_cov_matrix.transpose(2, 3)) + decorrelated_data = decorrelated_data.transpose(2, 3) + + # Reshape back to the original shape + decorrelated_data = decorrelated_data.view(num_samples, num_channels, height, width) + + return decorrelated_data.to(data.device) + +def get_low_frequency_noise(image: Tensor, threshold: float): + # Convert image to Fourier domain + fourier = torch.fft.fft2(image, dim=(-2, -1)) # Apply FFT along Height and Width dimensions + + # Compute the power spectrum + power_spectrum = torch.abs(fourier) ** 2 + + threshold = threshold ** 2 + + # Drop low-frequency components + mask = (power_spectrum < threshold).float() + filtered_fourier = fourier * mask + + # Inverse transform back to spatial domain + inverse_transformed = torch.fft.ifft2(filtered_fourier, dim=(-2, -1)) # Apply IFFT along Height and Width dimensions + + return inverse_transformed.real.to(image.device) + +def spectral_modulation(image: Tensor, modulation_multiplier: float, spectral_mod_percentile: float): # Reference implementation by Clybius, 2023 :tm::c::r: (jk idc who uses it :3) + # Convert image to Fourier domain + fourier = torch.fft.fft2(image, dim=(-2, -1)) # Apply FFT along Height and Width dimensions + + log_amp = torch.log(torch.sqrt(fourier.real ** 2 + fourier.imag ** 2)) + + quantile_low = torch.quantile( + log_amp.abs().flatten(2), + spectral_mod_percentile * 0.01, + dim = 2 + ).unsqueeze(-1).unsqueeze(-1).expand(log_amp.shape) + + quantile_high = torch.quantile( + log_amp.abs().flatten(2), + 1 - (spectral_mod_percentile * 0.01), + dim = 2 + ).unsqueeze(-1).unsqueeze(-1).expand(log_amp.shape) + + # Increase low-frequency components + mask_low = ((log_amp < quantile_low).float() + 1).clamp_(max=1.5) # If lower than low 5% quantile, set to 1.5, otherwise 1 + # Decrease high-frequency components + mask_high = ((log_amp < quantile_high).float()).clamp_(min=0.5) # If lower than high 5% quantile, set to 1, otherwise 0.5 + filtered_fourier = fourier * ((mask_low * mask_high) ** modulation_multiplier) # Effectively + + # Inverse transform back to spatial domain + inverse_transformed = torch.fft.ifft2(filtered_fourier, dim=(-2, -1)) # Apply IFFT along Height and Width dimensions + + return inverse_transformed.real.to(image.device) + +def spectral_modulation_soft(image: Tensor, modulation_multiplier: float, spectral_mod_percentile: float): # Modified for soft quantile adjustment using a novel:tm::c::r: method titled linalg. + # Convert image to Fourier domain + fourier = torch.fft.fft2(image, dim=(-2, -1)) # Apply FFT along Height and Width dimensions + + log_amp = torch.log(torch.sqrt(fourier.real ** 2 + fourier.imag ** 2)) + + quantile_low = torch.quantile( + log_amp.abs().flatten(2), + spectral_mod_percentile * 0.01, + dim = 2 + ).unsqueeze(-1).unsqueeze(-1).expand(log_amp.shape) + + quantile_high = torch.quantile( + log_amp.abs().flatten(2), + 1 - (spectral_mod_percentile * 0.01), + dim = 2 + ).unsqueeze(-1).unsqueeze(-1).expand(log_amp.shape) + + quantile_max = torch.quantile( + log_amp.abs().flatten(2), + 1, + dim = 2 + ).unsqueeze(-1).unsqueeze(-1).expand(log_amp.shape) + + # Decrease high-frequency components + mask_high = log_amp > quantile_high # If we're larger than 95th percentile + + additive_mult_high = torch.where( + mask_high, + 1 - ((log_amp - quantile_high) / (quantile_max - quantile_high)).clamp_(max=0.5), # (1) - (0-1), where 0 is 95th %ile and 1 is 100%ile + torch.tensor(1.0) + ) + + + # Increase low-frequency components + mask_low = log_amp < quantile_low + additive_mult_low = torch.where( + mask_low, + 1 + (1 - (log_amp / quantile_low)).clamp_(max=0.5), # (1) + (0-1), where 0 is 5th %ile and 1 is 0%ile + torch.tensor(1.0) + ) + + mask_mult = ((additive_mult_low * additive_mult_high) ** modulation_multiplier).clamp_(min=0.05, max=20) + #print(mask_mult) + filtered_fourier = fourier * mask_mult + + # Inverse transform back to spatial domain + inverse_transformed = torch.fft.ifft2(filtered_fourier, dim=(-2, -1)) # Apply IFFT along Height and Width dimensions + + return inverse_transformed.real.to(image.device) + +def pyramid_noise_like(x, discount=0.9, generator=None, rand_source=random): + b, c, w, h = x.shape # EDIT: w and h get over-written, rename for a different variant! + u = torch.nn.Upsample(size=(w, h), mode='nearest-exact') + noise = gen_like(torch.randn, x, generator=generator) + for i in range(10): + r = rand_source.random()*2+2 # Rather than always going 2x, + w, h = max(1, int(w/(r**i))), max(1, int(h/(r**i))) + noise += u(torch.randn(b, c, w, h, generator=generator).to(x)) * discount**i + if w==1 or h==1: break # Lowest resolution is 1x1 + return noise/noise.std() # Scaled back to roughly unit variance + +import math +def dyn_cfg_modifier(conditioning, unconditioning, method, cond_scale, time_mult): + match method: + case "dyncfg-halfcosine": + noise_pred = conditioning - unconditioning + + noise_pred_magnitude = (torch.linalg.vector_norm(noise_pred, dim=(1)) + 0.0000000001)[:,None] + + time = time_mult.item() + time_factor = -(math.cos(0.5 * time * math.pi) / 2) + 1 + noise_pred_timescaled_magnitude = (torch.linalg.vector_norm(noise_pred * time_factor, dim=(1)) + 0.0000000001)[:,None] + + noise_pred /= noise_pred_magnitude + noise_pred *= noise_pred_timescaled_magnitude + return noise_pred + case "dyncfg-halfcosine-mimic": + noise_pred = conditioning - unconditioning + + noise_pred_magnitude = (torch.linalg.vector_norm(noise_pred, dim=(1)) + 0.0000000001)[:,None] + + time = time_mult.item() + time_factor = -(math.cos(0.5 * time * math.pi) / 2) + 1 + + latent = noise_pred + + mimic_latent = noise_pred * time_factor + mimic_flattened = mimic_latent.flatten(2) + mimic_means = mimic_flattened.mean(dim=2).unsqueeze(2) + mimic_recentered = mimic_flattened - mimic_means + mimic_abs = mimic_recentered.abs() + mimic_max = mimic_abs.max(dim=2).values.unsqueeze(2) + + latent_flattened = latent.flatten(2) + latent_means = latent_flattened.mean(dim=2).unsqueeze(2) + latent_recentered = latent_flattened - latent_means + latent_abs = latent_recentered.abs() + latent_q = torch.quantile(latent_abs, 0.995, dim=2).unsqueeze(2) + s = torch.maximum(latent_q, mimic_max) + pred_clamped = noise_pred.flatten(2).clamp(-s, s) + pred_normalized = pred_clamped / s + pred_renorm = pred_normalized * mimic_max + pred_uncentered = pred_renorm + latent_means + noise_pred_degraded = pred_uncentered.unflatten(2, noise_pred.shape[2:]) + + noise_pred /= noise_pred_magnitude + + noise_pred_timescaled_magnitude = (torch.linalg.vector_norm(noise_pred_degraded, dim=(1)) + 0.0000000001)[:,None] + noise_pred *= noise_pred_timescaled_magnitude + return noise_pred + + +class ModelSamplerLatentMegaModifier: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "sharpness_multiplier": ("FLOAT", {"default": 0.0, "min": -100.0, "max": 100.0, "step": 0.1}), + "sharpness_method": (["anisotropic", "joint-anisotropic", "gaussian", "cas"], ), + "tonemap_multiplier": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step": 0.01}), + "tonemap_method": (["reinhard", "reinhard_perchannel", "arctan", "quantile", "gated", "cfg-mimic", "spatial-norm"], ), + "tonemap_percentile": ("FLOAT", {"default": 100.0, "min": 0.0, "max": 100.0, "step": 0.005}), + "contrast_multiplier": ("FLOAT", {"default": 0.0, "min": -100.0, "max": 100.0, "step": 0.1}), + "combat_method": (["subtract", "subtract_channels", "subtract_median", "sharpen"], ), + "combat_cfg_drift": ("FLOAT", {"default": 0.0, "min": -10.0, "max": 10.0, "step": 0.01}), + "rescale_cfg_phi": ("FLOAT", {"default": 0.0, "min": -10.0, "max": 10.0, "step": 0.01}), + "extra_noise_type": (["gaussian", "uniform", "perlin", "pink", "green", "pyramid"], ), + "extra_noise_method": (["add", "add_scaled", "speckle", "cads", "cads_rescaled", "cads_speckle", "cads_speckle_rescaled"], ), + "extra_noise_multiplier": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step": 0.1}), + "extra_noise_lowpass": ("INT", {"default": 100, "min": 0, "max": 1000, "step": 1}), + "divisive_norm_size": ("INT", {"default": 127, "min": 1, "max": 255, "step": 1}), + "divisive_norm_multiplier": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01}), + "spectral_mod_mode": (["hard_clamp", "soft_clamp"], ), + "spectral_mod_percentile": ("FLOAT", {"default": 5.0, "min": 0.0, "max": 50.0, "step": 0.01}), + "spectral_mod_multiplier": ("FLOAT", {"default": 0.0, "min": -15.0, "max": 15.0, "step": 0.01}), + "affect_uncond": (["None", "Sharpness"], ), + "dyn_cfg_augmentation": (["None", "dyncfg-halfcosine", "dyncfg-halfcosine-mimic"], ), + }, + "optional": { "seed": ("INT", {"min": 0, "max": 0xffffffffffffffff}) + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "mega_modify" + + CATEGORY = "clybNodes" + + def mega_modify(self, model, sharpness_multiplier, sharpness_method, tonemap_multiplier, tonemap_method, tonemap_percentile, contrast_multiplier, combat_method, combat_cfg_drift, rescale_cfg_phi, extra_noise_type, extra_noise_method, extra_noise_multiplier, extra_noise_lowpass, divisive_norm_size, divisive_norm_multiplier, spectral_mod_mode, spectral_mod_percentile, spectral_mod_multiplier, affect_uncond, dyn_cfg_augmentation, seed=None): + gen = None + rand = random + if seed is not None: + gen = torch.Generator(device='cpu') + rand = random.Random() + gen.manual_seed(seed) + rand.seed(seed) + + def modify_latent(args): + x_input = args["input"] + cond = args["cond"] + uncond = args["uncond"] + cond_scale = args["cond_scale"] + timestep = model.model.model_sampling.timestep(args["timestep"]) + sigma = args["sigma"] + sigma = sigma.view(sigma.shape[:1] + (1,) * (cond.ndim - 1)) + #print(model.model.model_sampling.timestep(timestep)) + + x = x_input / (sigma * sigma + 1.0) + cond = ((x - (x_input - cond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma) + uncond = ((x - (x_input - uncond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma) + + noise_pred = (cond - uncond) + + # Extra noise + if extra_noise_multiplier > 0: + match extra_noise_type: + case "gaussian": + extra_noise = gen_like(torch.randn, cond, generator=gen) + case "uniform": + extra_noise = (gen_like(torch.rand, cond, generator=gen) - 0.5) * 2 * 1.73 + case "perlin": + cond_size_0 = cond.size(dim=2) + cond_size_1 = cond.size(dim=3) + extra_noise = perlin_noise(grid_shape=(cond_size_0, cond_size_1), out_shape=(cond_size_0, cond_size_1), batch_size=4, generator=gen).to(cond.device).unsqueeze(0) + mean = torch.mean(extra_noise) + std = torch.std(extra_noise) + + extra_noise.sub_(mean).div_(std) + case "pink": + extra_noise = generate_1f_noise(cond, 2, extra_noise_multiplier, generator=gen).to(cond.device) + mean = torch.mean(extra_noise) + std = torch.std(extra_noise) + + extra_noise.sub_(mean).div_(std) + case "green": + cond_size_0 = cond.size(dim=2) + cond_size_1 = cond.size(dim=3) + extra_noise = green_noise(cond_size_0, cond_size_1, generator=gen).to(cond.device) + mean = torch.mean(extra_noise) + std = torch.std(extra_noise) + + extra_noise.sub_(mean).div_(std) + case "pyramid": + extra_noise = pyramid_noise_like(cond) + + if extra_noise_lowpass > 0: + extra_noise = get_low_frequency_noise(extra_noise, extra_noise_lowpass) + + alpha_noise = 1.0 - (timestep / 999.0)[:, None, None, None].clone() # Get alpha multiplier, lower alpha at high sigmas/high noise + alpha_noise *= 0.001 * extra_noise_multiplier # User-input and weaken the strength so we don't annihilate the latent. + match extra_noise_method: + case "add": + cond = cond + extra_noise * alpha_noise + uncond = uncond - extra_noise * alpha_noise + case "add_scaled": + cond = cond + train_difference(cond, extra_noise, cond) * alpha_noise + uncond = uncond - train_difference(uncond, extra_noise, uncond) * alpha_noise + case "speckle": + cond = cond + cond * extra_noise * alpha_noise + uncond = uncond - uncond * extra_noise * alpha_noise + case "cads": + cond = add_cads_custom_noise(cond, extra_noise, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, False) + uncond = add_cads_custom_noise(uncond, extra_noise, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, False) + case "cads_rescaled": + cond = add_cads_custom_noise(cond, extra_noise, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, True) + uncond = add_cads_custom_noise(uncond, extra_noise, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, True) + case "cads_speckle": + cond = add_cads_custom_noise(cond, extra_noise * cond, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, False) + uncond = add_cads_custom_noise(uncond, extra_noise * uncond, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, False) + case "cads_speckle_rescaled": + cond = add_cads_custom_noise(cond, extra_noise * cond, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, True) + uncond = add_cads_custom_noise(uncond, extra_noise * uncond, timestep, 0.6, 0.9, extra_noise_multiplier / 100., 1, True) + case _: + print("Haven't heard of a noise method named like that before... (Couldn't find method)") + + if sharpness_multiplier > 0.0 or sharpness_multiplier < 0.0: + match sharpness_method: + case "anisotropic": + degrade_func = bilateral_blur + case "joint-anisotropic": + degrade_func = lambda img: joint_bilateral_blur(img, (img - torch.mean(img, dim=(1, 2, 3), keepdim=True)) / torch.std(img, dim=(1, 2, 3), keepdim=True), 13, 3.0, 3.0, "reflect", "l1") + case "gaussian": + degrade_func = gaussian_filter_2d + case "cas": + degrade_func = lambda image: contrast_adaptive_sharpening(image, amount=sigma.clamp(max=1.00).item()) + case _: + print("For some reason, the sharpness filter could not be found.") + # Sharpness + alpha = 1.0 - (timestep / 999.0)[:, None, None, None].clone() # Get alpha multiplier, lower alpha at high sigmas/high noise + alpha *= 0.001 * sharpness_multiplier # User-input and weaken the strength so we don't annihilate the latent. + cond = degrade_func(cond) * alpha + cond * (1.0 - alpha) # Mix the modified latent with the existing latent by the alpha + if affect_uncond == "Sharpness": + uncond = degrade_func(uncond) * alpha + uncond * (1.0 - alpha) + + time_mult = 1.0 - (timestep / 999.0)[:, None, None, None].clone() + noise_pred_degraded = (cond - uncond) if dyn_cfg_augmentation == "None" else dyn_cfg_modifier(cond, uncond, dyn_cfg_augmentation, cond_scale, time_mult) # New noise pred + + # After this point, we use `noise_pred_degraded` instead of just `cond` for the final set of calculations + + # Tonemap noise + if tonemap_multiplier == 0: + new_magnitude = 1.0 + else: + match tonemap_method: + case "reinhard": + noise_pred_vector_magnitude = (torch.linalg.vector_norm(noise_pred_degraded, dim=(1)) + 0.0000000001)[:,None] + noise_pred_degraded /= noise_pred_vector_magnitude + + mean = torch.mean(noise_pred_vector_magnitude, dim=(1,2,3), keepdim=True) + std = torch.std(noise_pred_vector_magnitude, dim=(1,2,3), keepdim=True) + + top = (std * 3 * (100 / tonemap_percentile) + mean) * tonemap_multiplier + + noise_pred_vector_magnitude *= (1.0 / top) + new_magnitude = noise_pred_vector_magnitude / (noise_pred_vector_magnitude + 1.0) + new_magnitude *= top + + noise_pred_degraded *= new_magnitude + case "reinhard_perchannel": # Testing the flatten strategy + flattened = noise_pred_degraded.flatten(2) + noise_pred_vector_magnitude = (torch.linalg.vector_norm(flattened, dim=(2), keepdim=True) + 0.0000000001) + flattened /= noise_pred_vector_magnitude + + mean = torch.mean(noise_pred_vector_magnitude, dim=(2), keepdim=True) + + top = (3 * (100 / tonemap_percentile) + mean) * tonemap_multiplier + + noise_pred_vector_magnitude *= (1.0 / top) + + new_magnitude = noise_pred_vector_magnitude / (noise_pred_vector_magnitude + 1.0) + new_magnitude *= top + + flattened *= new_magnitude + noise_pred_degraded = flattened.unflatten(2, noise_pred_degraded.shape[2:]) + case "arctan": + noise_pred_vector_magnitude = (torch.linalg.vector_norm(noise_pred_degraded, dim=(1)) + 0.0000000001)[:,None] + noise_pred_degraded /= noise_pred_vector_magnitude + + noise_pred_degraded = (torch.arctan(noise_pred_degraded * tonemap_multiplier) * (1 / tonemap_multiplier)) + (noise_pred_degraded * (100 - tonemap_percentile) / 100) + + noise_pred_degraded *= noise_pred_vector_magnitude + case "quantile": + s: FloatTensor = torch.quantile( + (uncond + noise_pred_degraded * cond_scale).flatten(start_dim=1).abs(), + tonemap_percentile / 100, + dim = -1 + ) * tonemap_multiplier + s.clamp_(min = 1.) + s = s.reshape(*s.shape, 1, 1, 1) + noise_pred_degraded = noise_pred_degraded.clamp(-s, s) / s + case "gated": # https://birchlabs.co.uk/machine-learning#dynamic-thresholding-latents so based,.,.,...., + latent_scale = model.model.latent_format.scale_factor + + latent = uncond + noise_pred_degraded * cond_scale # Get full latent from CFG formula + latent /= latent_scale # Divide full CFG by latent scale (~0.13 for sdxl) + flattened = latent.flatten(2) + means = flattened.mean(dim=2).unsqueeze(2) + centered_magnitudes = (flattened - means).abs().max() # Get highest magnitude of full CFG + + flattened_pred = (noise_pred_degraded / latent_scale).flatten(2) + + floor = 3.0560 + ceil = 42. * tonemap_multiplier # as is the answer to life, unless you modify the multiplier cuz u aint a believer in life + + + thresholded_latent = dyn_thresh_gate(flattened_pred, centered_magnitudes, tonemap_percentile / 100., floor, ceil) # Threshold if passes ceil + thresholded_latent = thresholded_latent.unflatten(2, noise_pred_degraded.shape[2:]) + noise_pred_degraded = thresholded_latent * latent_scale # Rescale by latent + case "cfg-mimic": + latent = noise_pred_degraded + + mimic_latent = noise_pred_degraded * tonemap_multiplier + mimic_flattened = mimic_latent.flatten(2) + mimic_means = mimic_flattened.mean(dim=2).unsqueeze(2) + mimic_recentered = mimic_flattened - mimic_means + mimic_abs = mimic_recentered.abs() + mimic_max = mimic_abs.max(dim=2).values.unsqueeze(2) + + latent_flattened = latent.flatten(2) + latent_means = latent_flattened.mean(dim=2).unsqueeze(2) + latent_recentered = latent_flattened - latent_means + latent_abs = latent_recentered.abs() + latent_q = torch.quantile(latent_abs, tonemap_percentile / 100., dim=2).unsqueeze(2) + s = torch.maximum(latent_q, mimic_max) + pred_clamped = noise_pred_degraded.flatten(2).clamp(-s, s) + pred_normalized = pred_clamped / s + pred_renorm = pred_normalized * mimic_max + pred_uncentered = pred_renorm + mimic_means # Personal choice to re-mean from the mimic here... should be latent_means. + noise_pred_degraded = pred_uncentered.unflatten(2, noise_pred_degraded.shape[2:]) + case "spatial-norm": + #time = (1.0 - (timestep / 999.0)[:, None, None, None].clone().item()) + #time = -(math.cos(time * math.pi) / (3)) + (2/3) # 0.33333 to 1.0, half cosine + noise_pred_degraded = spatial_norm_chw_thresholding(noise_pred_degraded, tonemap_multiplier / 2 / cond_scale) + case _: + print("Could not tonemap, for the method was not found.") + + # Spectral Modification + if spectral_mod_multiplier > 0 or spectral_mod_multiplier < 0: + #alpha = 1. - (timestep / 999.0)[:, None, None, None].clone() # Get alpha multiplier, lower alpha at high sigmas/high noise + #alpha = spectral_mod_multiplier# User-input and weaken the strength so we don't annihilate the latent. + match spectral_mod_mode: + case "hard_clamp": + modulation_func = spectral_modulation + case "soft_clamp": + modulation_func = spectral_modulation_soft + modulation_diff = modulation_func(noise_pred_degraded, spectral_mod_multiplier, spectral_mod_percentile) - noise_pred_degraded + noise_pred_degraded += modulation_diff + + if contrast_multiplier > 0 or contrast_multiplier < 0: + contrast_func = contrast + # Contrast, after tonemapping, to ensure user-set contrast is expected to behave similarly across tonemapping settings + alpha = 1.0 - (timestep / 999.0)[:, None, None, None].clone() + alpha *= 0.001 * contrast_multiplier + noise_pred_degraded = contrast_func(noise_pred_degraded) * alpha + (noise_pred_degraded) * (1.0 - alpha) # Temporary fix for contrast is to add the input? Maybe? It just doesn't work like before... + + # Rescale CFG + if rescale_cfg_phi == 0: + x_final = uncond + noise_pred_degraded * cond_scale + else: + x_cfg = uncond + noise_pred_degraded * cond_scale + ro_pos = torch.std(cond, dim=(1,2,3), keepdim=True) + ro_cfg = torch.std(x_cfg, dim=(1,2,3), keepdim=True) + + x_rescaled = x_cfg * (ro_pos / ro_cfg) + x_final = rescale_cfg_phi * x_rescaled + (1.0 - rescale_cfg_phi) * x_cfg + + if combat_cfg_drift > 0 or combat_cfg_drift < 0: + alpha = (1. - (timestep / 999.0)[:, None, None, None].clone()) + alpha ** 0.025 # Alpha might as well be 1, but we want to protect the first steps (?). + alpha = alpha.clamp_(max=1) + match combat_method: + case "subtract": + combat_drift_func = center_latent_perchannel + alpha *= combat_cfg_drift + case "subtract_channels": + combat_drift_func = center_0channel + alpha *= combat_cfg_drift + case "subtract_median": + combat_drift_func = center_latent_median + alpha *= combat_cfg_drift + case "sharpen": + combat_drift_func = channel_sharpen + alpha *= combat_cfg_drift + x_final = combat_drift_func(x_final) * alpha + x_final * (1.0 - alpha) # Mix the modified latent with the existing latent by the alpha + + if divisive_norm_multiplier > 0: + alpha = 1. - (timestep / 999.0)[:, None, None, None].clone() + alpha ** 0.025 # Alpha might as well be 1, but we want to protect the beginning steps (?). + alpha *= divisive_norm_multiplier + high_noise = divisive_normalization(x_final, (divisive_norm_size * 2) + 1) + x_final = high_noise * alpha + x_final * (1.0 - alpha) + + + return x_input - (x - x_final * sigma / (sigma * sigma + 1.0) ** 0.5) # General formula for CFG. uncond + (cond - uncond) * cond_scale + + m = model.clone() + m.set_model_sampler_cfg_function(modify_latent) + return (m, ) \ No newline at end of file diff --git a/extensions-builtin/sd_forge_latent_modifier/scripts/forge_latent_modifier.py b/extensions-builtin/sd_forge_latent_modifier/scripts/forge_latent_modifier.py new file mode 100644 index 00000000..dd641399 --- /dev/null +++ b/extensions-builtin/sd_forge_latent_modifier/scripts/forge_latent_modifier.py @@ -0,0 +1,104 @@ +import gradio as gr +from modules import scripts + +from lib_latent_modifier.sampler_mega_modifier import ModelSamplerLatentMegaModifier + +opModelSamplerLatentMegaModifier = ModelSamplerLatentMegaModifier().mega_modify + + +class LatentModifierForForge(scripts.Script): + def title(self): + return "LatentModifier Integrated" + + def show(self, is_img2img): + # make this extension visible in both txt2img and img2img tab. + return scripts.AlwaysVisible + + def ui(self, *args, **kwargs): + with gr.Accordion(open=False, label=self.title()): + enabled = gr.Checkbox(label='Enabled', value=False) + sharpness_multiplier = gr.Slider(label='Sharpness Multiplier', minimum=-100.0, maximum=100.0, step=0.1, + value=0.0) + sharpness_method = gr.Radio(label='Sharpness Method', + choices=['anisotropic', 'joint-anisotropic', 'gaussian', 'cas'], + value='anisotropic') + tonemap_multiplier = gr.Slider(label='Tonemap Multiplier', minimum=0.0, maximum=100.0, step=0.01, value=0.0) + tonemap_method = gr.Radio(label='Tonemap Method', + choices=['reinhard', 'reinhard_perchannel', 'arctan', 'quantile', 'gated', + 'cfg-mimic', 'spatial-norm'], value='reinhard') + tonemap_percentile = gr.Slider(label='Tonemap Percentile', minimum=0.0, maximum=100.0, step=0.005, + value=100.0) + contrast_multiplier = gr.Slider(label='Contrast Multiplier', minimum=-100.0, maximum=100.0, step=0.1, + value=0.0) + combat_method = gr.Radio(label='Combat Method', + choices=['subtract', 'subtract_channels', 'subtract_median', 'sharpen'], + value='subtract') + combat_cfg_drift = gr.Slider(label='Combat Cfg Drift', minimum=-10.0, maximum=10.0, step=0.01, value=0.0) + rescale_cfg_phi = gr.Slider(label='Rescale Cfg Phi', minimum=-10.0, maximum=10.0, step=0.01, value=0.0) + extra_noise_type = gr.Radio(label='Extra Noise Type', + choices=['gaussian', 'uniform', 'perlin', 'pink', 'green', 'pyramid'], + value='gaussian') + extra_noise_method = gr.Radio(label='Extra Noise Method', + choices=['add', 'add_scaled', 'speckle', 'cads', 'cads_rescaled', + 'cads_speckle', 'cads_speckle_rescaled'], value='add') + extra_noise_multiplier = gr.Slider(label='Extra Noise Multiplier', minimum=0.0, maximum=100.0, step=0.1, + value=0.0) + extra_noise_lowpass = gr.Slider(label='Extra Noise Lowpass', minimum=0, maximum=1000, step=1, value=100) + divisive_norm_size = gr.Slider(label='Divisive Norm Size', minimum=1, maximum=255, step=1, value=127) + divisive_norm_multiplier = gr.Slider(label='Divisive Norm Multiplier', minimum=0.0, maximum=1.0, step=0.01, + value=0.0) + spectral_mod_mode = gr.Radio(label='Spectral Mod Mode', choices=['hard_clamp', 'soft_clamp'], + value='hard_clamp') + spectral_mod_percentile = gr.Slider(label='Spectral Mod Percentile', minimum=0.0, maximum=50.0, step=0.01, + value=5.0) + spectral_mod_multiplier = gr.Slider(label='Spectral Mod Multiplier', minimum=-15.0, maximum=15.0, step=0.01, + value=0.0) + affect_uncond = gr.Radio(label='Affect Uncond', choices=['None', 'Sharpness'], value='None') + dyn_cfg_augmentation = gr.Radio(label='Dyn Cfg Augmentation', + choices=['None', 'dyncfg-halfcosine', 'dyncfg-halfcosine-mimic'], + value='None') + + return enabled, sharpness_multiplier, sharpness_method, tonemap_multiplier, tonemap_method, tonemap_percentile, contrast_multiplier, combat_method, combat_cfg_drift, rescale_cfg_phi, extra_noise_type, extra_noise_method, extra_noise_multiplier, extra_noise_lowpass, divisive_norm_size, divisive_norm_multiplier, spectral_mod_mode, spectral_mod_percentile, spectral_mod_multiplier, affect_uncond, dyn_cfg_augmentation + + def process_before_every_sampling(self, p, *script_args, **kwargs): + # This will be called before every sampling. + # If you use highres fix, this will be called twice. + + enabled, sharpness_multiplier, sharpness_method, tonemap_multiplier, tonemap_method, tonemap_percentile, contrast_multiplier, combat_method, combat_cfg_drift, rescale_cfg_phi, extra_noise_type, extra_noise_method, extra_noise_multiplier, extra_noise_lowpass, divisive_norm_size, divisive_norm_multiplier, spectral_mod_mode, spectral_mod_percentile, spectral_mod_multiplier, affect_uncond, dyn_cfg_augmentation = script_args + + if not enabled: + return + + unet = p.sd_model.forge_objects.unet + + unet = opModelSamplerLatentMegaModifier(unet, sharpness_multiplier, sharpness_method, tonemap_multiplier, tonemap_method, tonemap_percentile, contrast_multiplier, combat_method, combat_cfg_drift, rescale_cfg_phi, extra_noise_type, extra_noise_method, extra_noise_multiplier, extra_noise_lowpass, divisive_norm_size, divisive_norm_multiplier, spectral_mod_mode, spectral_mod_percentile, spectral_mod_multiplier, affect_uncond, dyn_cfg_augmentation, seed=p.seeds[0])[0] + + p.sd_model.forge_objects.unet = unet + + # Below codes will add some logs to the texts below the image outputs on UI. + # The extra_generation_params does not influence results. + p.extra_generation_params.update(dict( + latent_modifier_enabled=enabled, + latent_modifier_sharpness_multiplier=sharpness_multiplier, + latent_modifier_sharpness_method=sharpness_method, + latent_modifier_tonemap_multiplier=tonemap_multiplier, + latent_modifier_tonemap_method=tonemap_method, + latent_modifier_tonemap_percentile=tonemap_percentile, + latent_modifier_contrast_multiplier=contrast_multiplier, + latent_modifier_combat_method=combat_method, + latent_modifier_combat_cfg_drift=combat_cfg_drift, + latent_modifier_rescale_cfg_phi=rescale_cfg_phi, + latent_modifier_extra_noise_type=extra_noise_type, + latent_modifier_extra_noise_method=extra_noise_method, + latent_modifier_extra_noise_multiplier=extra_noise_multiplier, + latent_modifier_extra_noise_lowpass=extra_noise_lowpass, + latent_modifier_divisive_norm_size=divisive_norm_size, + latent_modifier_divisive_norm_multiplier=divisive_norm_multiplier, + latent_modifier_spectral_mod_mode=spectral_mod_mode, + latent_modifier_spectral_mod_percentile=spectral_mod_percentile, + latent_modifier_spectral_mod_multiplier=spectral_mod_multiplier, + latent_modifier_affect_uncond=affect_uncond, + latent_modifier_dyn_cfg_augmentation=dyn_cfg_augmentation, + )) + + return diff --git a/modules_forge/gradio_compile.py b/modules_forge/gradio_compile.py index 739ae987..5fa16525 100644 --- a/modules_forge/gradio_compile.py +++ b/modules_forge/gradio_compile.py @@ -43,7 +43,16 @@ def gradio_compile(items, prefix): else: print('error ' + str(t)) - return names + return ['enabled'] + names + + +def print_info_text(name_list, prefix): + print(', '.join(name_list)) + print('p.extra_generation_params.update(dict(') + for n in name_list: + print(prefix + '_' + n + ' = ' + n + ', ') + print(')') + return # from modules_forge.gradio_compile import gradio_compile @@ -52,3 +61,4 @@ def gradio_compile(items, prefix): # ps += gradio_compile(KSampler.INPUT_TYPES(), prefix='sampling') # ps += gradio_compile(VideoLinearCFGGuidance.INPUT_TYPES(), prefix='guidance') # print(', '.join(ps)) +# print_info_text(ps, '123')