Update README.md

Fix for latest ComfyUI
version 1.0.3
2026-04-29 18:30:56 +03:00 · 2026-03-15 17:02:02 +02:00 · 2026-02-04 16:55:13 +02:00 · 2026-02-04 16:53:20 +02:00 · 2026-02-04 16:49:40 +02:00 · 2026-02-04 16:48:18 +02:00
23 changed files with 2804 additions and 902 deletions
@@ -0,0 +1,25 @@
 name: Publish to Comfy registry
 on:
  workflow_dispatch:
  push:
    branches:
      - main
    paths:
      - "pyproject.toml"
 permissions:
  issues: write
 jobs:
  publish-node:
    name: Publish Custom Node to registry
    runs-on: ubuntu-latest
    if: ${{ github.repository_owner == 'kijai' }}
    steps:
      - name: Check out code
        uses: actions/checkout@v4
      - name: Publish Custom Node
        uses: Comfy-Org/publish-node-action@v1
        with:
          ## Add your own personal access token to your Github Repository secrets and reference it here.
          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
@@ -1,674 +1,13 @@
-GNU GENERAL PUBLIC LICENSE
+License
                       Version 3, 29 June 2007
- Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+Copyright (c) 2024 XPixel Group, Especially the author team of SUPIR.
 Everyone is permitted to copy and distribute verbatim copies
 of this license document, but changing it is not allowed.
-                            Preamble
+The SUPIR ("Software") is made available for use, reproduction, and distribution strictly for non-commercial purposes. For the purposes of this declaration, "non-commercial" is defined as not primarily intended for or directed towards commercial advantage or monetary compensation.
-  The GNU General Public License is a free, copyleft license for
+By using, reproducing, or distributing the Software, you agree to abide by this restriction and not to use the Software for any commercial purposes without obtaining prior written permission from Dr. Jinjin Gu.
 software and other kinds of works.
-  The licenses for most software and other practical works are designed
+This declaration does not in any way limit the rights under any open source license that may apply to the Software; it solely adds a condition that the Software shall not be used for commercial purposes.
 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
+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.
 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
+For inquiries or to obtain permission for commercial use, please contact Dr. Jinjin Gu (jinjin.gu@suppixel.ai).
 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.
    <one line to give the program's name and a brief idea of what it does.>
    Copyright (C) <year>  <name of author>
    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 <https://www.gnu.org/licenses/>.
 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:
    <program>  Copyright (C) <year>  <name of author>
    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
 <https://www.gnu.org/licenses/>.
  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
 <https://www.gnu.org/licenses/why-not-lgpl.html>.
@@ -1,7 +1,29 @@
 # ComfyUI SUPIR upscaler wrapper node
-## WORK IN PROGRESS
+# FINAL update
-![image](https://github.com/kijai/ComfyUI-SUPIR/assets/40791699/887898d3-afe5-45d1-be08-50f6620b70eb)
+
 SUPIR is now available to use in ComfyUI core after this PR: 
 https://github.com/Comfy-Org/ComfyUI/pull/13250
 These nodes won't be updated beyond simple breaking bugfixes.
 # UPDATE3:
 Pruned models in safetensors format now available here:
 https://huggingface.co/Kijai/SUPIR_pruned/tree/main
 # UPDATE2:
 ![image](https://github.com/kijai/ComfyUI-SUPIR/assets/40791699/65baec3e-cb4a-4eec-8d45-2b08157b1e86)
 Added a better way to load the SDXL model, which also allows using LoRAs. The old node will remain for now to not break old workflows, and it is dubbed Legacy along with the single node, as I do not want to maintain those.
 # UPDATE:
 As I have learned a lot with this project, I have now separated the single node to multiple nodes that make more sense to use in ComfyUI, and makes it clearer how SUPIR works. This is still a wrapper, though the whole thing has deviated from the original with much wider hardware support, more efficient model loading, far less memory usage and more sampler options. Here's a quick example (workflow is included) of using a Ligntning model, quality suffers then but it's very fast and I recommend starting with it as faster sampling makes it a lot easier to learn what the settings do.
 Under the hood SUPIR is SDXL img2img pipeline, the biggest custom part being their ControlNet. What they call "first stage" is a denoising process using their special "denoise encoder" VAE. This is not to be confused with the Gradio demo's "first stage" that's labeled as such for the Llava preprocessing, the Gradio "Stage2" still runs the denoising process anyway. This can be fully skipped with the nodes, or replaced with any other preprocessing node such as a model upscaler or anything you want.
 https://github.com/kijai/ComfyUI-SUPIR/assets/40791699/5cae2a24-d425-462c-b89d-df7dcf01595c
 # Installing
 Either manager and install from git, or clone this repo to custom_nodes and run:
@@ -25,6 +47,10 @@ I have not included llava in this, but you can input any captions to the node an
 Memory requirements are directly related to the input image resolution, the "scale_by" in the node simply scales the input, you can leave it at 1.0 and size your input with any other node as well. In my testing I was able to run 512x512 to 1024x1024 with a 10GB 3080 GPU, and other tests on 24GB GPU to up 3072x3072. System RAM requirements are also hefty, don't know numbers but I would guess under 32GB is going to have issues, tested with 64GB.
 ## Updates: 
 - fp8 seems to work fine for the unet, I was able to do 512p to 2048 with under 10GB VRAM used. For the VAE it seems to cause artifacts, I recommend using tiled_vae instead.
 - CLIP models are no longer needed separately, instead they are loaded from your selected SDXL checkpoint
 ______
 Mirror for the models: https://huggingface.co/camenduru/SUPIR/tree/main
 # Tests
@@ -7,6 +7,7 @@ import random
 from ...SUPIR.utils.colorfix import wavelet_reconstruction, adaptive_instance_normalization
 from pytorch_lightning import seed_everything
 from ...SUPIR.utils.tilevae import VAEHook
 from ...SUPIR.util import convert_dtype
 from contextlib import nullcontext
 import comfy.model_management
@@ -20,23 +21,8 @@ class SUPIRModel(DiffusionEngine):
        self.first_stage_model.denoise_encoder = copy.deepcopy(self.first_stage_model.encoder)
        self.sampler_config = kwargs['sampler_config']
-        assert (ae_dtype in ['fp32', 'fp16', 'bf16']) and (diffusion_dtype in ['fp32', 'fp16', 'bf16'])
+        self.ae_dtype = convert_dtype(ae_dtype)
-        if ae_dtype == 'fp32':
+        self.model.dtype = convert_dtype(diffusion_dtype)
            ae_dtype = torch.float32
        elif ae_dtype == 'fp16':
            raise RuntimeError('fp16 cause NaN in AE')
        elif ae_dtype == 'bf16':
            ae_dtype = torch.bfloat16
        if diffusion_dtype == 'fp32':
            diffusion_dtype = torch.float32
        elif diffusion_dtype == 'fp16':
            diffusion_dtype = torch.float16
        elif diffusion_dtype == 'bf16':
            diffusion_dtype = torch.bfloat16
        self.ae_dtype = ae_dtype
        self.model.dtype = diffusion_dtype
        self.p_p = p_p
        self.n_p = n_p
@@ -72,7 +58,6 @@ class SUPIRModel(DiffusionEngine):
    @torch.no_grad()
    def decode_first_stage(self, z):
        z = 1.0 / self.scale_factor * z
        #with torch.autocast(device, dtype=self.ae_dtype):
        autocast_condition = (self.ae_dtype == torch.float16 or self.ae_dtype == torch.bfloat16) and not comfy.model_management.is_device_mps(device)
        with torch.autocast(comfy.model_management.get_autocast_device(device), dtype=self.ae_dtype) if autocast_condition else nullcontext():
            out = self.first_stage_model.decode(z)
@@ -120,29 +105,51 @@ class SUPIRModel(DiffusionEngine):
        self.sampler_config.params.s_noise = s_noise
        self.sampler = instantiate_from_config(self.sampler_config)
        print("Sampler: ", self.sampler_config.target)
        print("sampler_config: ", self.sampler_config.params)
        if seed == -1:
            seed = random.randint(0, 65535)
        seed_everything(seed)
        self.model.to('cpu')
        self.conditioner.to('cpu')
        # stage 1: encode/decode/encode
        self.first_stage_model.to(device)
        _z = self.encode_first_stage_with_denoise(x, use_sample=False)
        x_stage1 = self.decode_first_stage(_z)
        z_stage1 = self.encode_first_stage(x_stage1)
        self.first_stage_model.to('cpu')
        #conditioning
        self.conditioner.to(device)
        c, uc = self.prepare_condition(_z, p, p_p, n_p, N)
        self.conditioner.to('cpu')
        denoiser = lambda input, sigma, c, control_scale: self.denoiser(
            self.model, input, sigma, c, control_scale, **kwargs
        )
        noised_z = torch.randn_like(_z).to(_z.device)
        comfy.model_management.soft_empty_cache()
        #sampling
        self.model.diffusion_model.to(device)
        self.model.control_model.to(device)
        self.denoiser.to(device)
        _samples = self.sampler(denoiser, noised_z, cond=c, uc=uc, x_center=z_stage1, control_scale=control_scale,
                                use_linear_control_scale=use_linear_control_scale, control_scale_start=control_scale_start)
        self.model.diffusion_model.to('cpu')
        self.model.control_model.to('cpu')
        #decoding
        self.first_stage_model.to(device)
        samples = self.decode_first_stage(_samples)
        self.first_stage_model.to('cpu')
        if color_fix_type == 'Wavelet':
            samples = wavelet_reconstruction(samples, x_stage1)
        elif color_fix_type == 'AdaIn':
@@ -193,19 +200,4 @@ class SUPIRModel(DiffusionEngine):
                    else:
                        _c, _ = self.conditioner.get_unconditional_conditioning(batch, None)
                c.append(_c)
-        return c, uc
+        return c, uc
 # if __name__ == '__main__':
 #     from SUPIR.util import create_model, load_state_dict
 #     model = create_model('../../options/dev/SUPIR_paper_version.yaml')
 #     SDXL_CKPT = '/opt/data/private/AIGC_pretrain/SDXL_cache/sd_xl_base_1.0_0.9vae.safetensors'
 #     SUPIR_CKPT = '/opt/data/private/AIGC_pretrain/SUPIR_cache/SUPIR-paper.ckpt'
 #     model.load_state_dict(load_state_dict(SDXL_CKPT), strict=False)
 #     model.load_state_dict(load_state_dict(SUPIR_CKPT), strict=False)
 #     model = model.cuda()
 #     x = torch.randn(1, 3, 512, 512).cuda()
 #     p = ['a professional, detailed, high-quality photo']
 #     samples = model.batchify_sample(x, p, num_steps=50, restoration_scale=4.0, s_churn=0, cfg_scale=4.0, seed=-1, num_samples=1)
@@ -0,0 +1,11 @@
 from ...sgm.models.diffusion import DiffusionEngine
 from ...sgm.util import instantiate_from_config
 import copy
 class SUPIRModel(DiffusionEngine):
    def __init__(self, control_stage_config, ae_dtype='fp32', diffusion_dtype='fp32', p_p='', n_p='', *args, **kwargs):
        super().__init__(*args, **kwargs)
        control_model = instantiate_from_config(control_stage_config)
        self.model.load_control_model(control_model)
        self.first_stage_model.denoise_encoder = copy.deepcopy(self.first_stage_model.encoder)
        self.sampler_config = kwargs['sampler_config']
@@ -27,6 +27,9 @@ from functools import partial
 import comfy.model_management
 device = comfy.model_management.get_torch_device()
 import comfy.ops
 ops = comfy.ops.manual_cast
 try:
    import xformers
    import xformers.ops
@@ -77,13 +80,13 @@ class ZeroSFT(nn.Module):
        nhidden = 128
        self.mlp_shared = nn.Sequential(
-            nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
+            ops.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
            nn.SiLU()
        )
-        self.zero_mul = zero_module(nn.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw))
+        self.zero_mul = zero_module(ops.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw))
-        self.zero_add = zero_module(nn.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw))
+        self.zero_add = zero_module(ops.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw))
-        # self.zero_mul = nn.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw)
+        # self.zero_mul = ops.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw)
-        # self.zero_add = nn.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw)
+        # self.zero_add = ops.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw)
        self.zero_conv = zero_module(conv_nd(2, label_nc, norm_nc, 1, 1, 0))
        self.pre_concat = bool(concat_channels != 0)
@@ -296,7 +299,7 @@ class GLVControl(nn.Module):
                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
            elif self.num_classes == "continuous":
                print("setting up linear c_adm embedding layer")
-                self.label_emb = nn.Linear(1, time_embed_dim)
+                self.label_emb = ops.Linear(1, time_embed_dim)
            elif self.num_classes == "timestep":
                self.label_emb = checkpoint_wrapper_fn(
                    nn.Sequential(
@@ -66,20 +66,20 @@ import torch
 import torch.version
 import torch.nn.functional as F
 from einops import rearrange
 #from diffusers.utils.import_utils import is_xformers_available
 #import SUPIR.utils.devices as devices
 import comfy.model_management
 device = comfy.model_management.get_torch_device()
-try:
+if comfy.model_management.XFORMERS_IS_AVAILABLE:
-    import xformers
+    try:
-    import xformers.ops
+        import xformers
-    XFORMERS_IS_AVAILABLE = True
+        import xformers.ops
-except:
+        XFORMERS_IS_AVAILABLE = True
    except:
        XFORMERS_IS_AVAILABLE = False
        print("no module 'xformers'. Processing without...")
 else:
    XFORMERS_IS_AVAILABLE = False
    print("no module 'xformers'. Processing without...")
 sd_flag = True
@@ -368,8 +368,14 @@ def attn2task(task_queue, net):
        task_queue.append(('pre_norm', net.norm))
        if XFORMERS_IS_AVAILABLE:
            # task_queue.append(('attn', lambda x, net=net: attn_forward_new_xformers(net, x)))
-            task_queue.append(
+            # Wrap xformer call with fallback to standard attention on NotImplementedError
-                ('attn', lambda x, net=net: xformer_attn_forward(net, x)))
+            # (e.g., new GPU architectures or unsupported dimensions)
            def xformer_with_fallback(x, net=net):
                try:
                    return xformer_attn_forward(net, x)
                except NotImplementedError:
                    return attn_forward(net, x)
            task_queue.append(('attn', xformer_with_fallback))
        elif hasattr(F, "scaled_dot_product_attention"):
            task_queue.append(('attn', lambda x, net=net: attn_forward(net, x)))
            #task_queue.append(('attn', lambda x, net=net: attn_forward_new_pt2_0(net, x)))
@@ -895,7 +901,7 @@ class VAEHook:
        # Task queue execution
        pbar = tqdm(total=num_tiles * len(task_queues[0]), desc=f"[Tiled VAE]: Executing {'Decoder' if is_decoder else 'Encoder'} Task Queue: ")
-
+        pbar_comfy = comfy.utils.ProgressBar(num_tiles * len(task_queues[0]))
        # execute the task back and forth when switch tiles so that we always
        # keep one tile on the GPU to reduce unnecessary data transfer
        forward = True
@@ -937,6 +943,7 @@ class VAEHook:
                        tile = task[1](tile)
                        #print(tiles[i].shape, tile.shape, task)
                    pbar.update(1)
                    pbar_comfy.update(1)
                if interrupted: break
@@ -1,3 +1,28 @@
-from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+from .nodes import SUPIR_Upscale
 from .nodes_v2 import SUPIR_sample, SUPIR_model_loader, SUPIR_first_stage, SUPIR_encode, SUPIR_decode, SUPIR_conditioner, SUPIR_tiles, SUPIR_model_loader_v2, SUPIR_model_loader_v2_clip
 NODE_CLASS_MAPPINGS = {
    "SUPIR_Upscale": SUPIR_Upscale,
    "SUPIR_sample": SUPIR_sample,
    "SUPIR_model_loader": SUPIR_model_loader,
    "SUPIR_first_stage": SUPIR_first_stage,
    "SUPIR_encode": SUPIR_encode,
    "SUPIR_decode": SUPIR_decode,
    "SUPIR_conditioner": SUPIR_conditioner,
    "SUPIR_tiles": SUPIR_tiles,
    "SUPIR_model_loader_v2": SUPIR_model_loader_v2,
    "SUPIR_model_loader_v2_clip": SUPIR_model_loader_v2_clip
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
    "SUPIR_Upscale": "SUPIR Upscale (Legacy)",
    "SUPIR_sample": "SUPIR Sampler",
    "SUPIR_model_loader": "SUPIR Model Loader (Legacy)",
    "SUPIR_first_stage": "SUPIR First Stage (Denoiser)",
    "SUPIR_encode": "SUPIR Encode",
    "SUPIR_decode": "SUPIR Decode",
    "SUPIR_conditioner": "SUPIR Conditioner",
    "SUPIR_tiles": "SUPIR Tiles Preview",
    "SUPIR_model_loader_v2": "SUPIR Model Loader (v2)",
    "SUPIR_model_loader_v2_clip": "SUPIR Model Loader (v2) (Clip)"
 }
 __all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
@@ -1,8 +1,6 @@
 import os
 import torch
 import torch.nn as nn
 from torch.nn import functional as F
 from contextlib import nullcontext
 from omegaconf import OmegaConf
 import comfy.utils
 import comfy.model_management as mm
@@ -13,6 +11,7 @@ import torch.cuda
 from .sgm.util import instantiate_from_config
 from .SUPIR.util import convert_dtype, load_state_dict
 import open_clip
 from contextlib import contextmanager
 from transformers import (
    CLIPTextModel,
@@ -22,24 +21,24 @@ from transformers import (
 )
 script_directory = os.path.dirname(os.path.abspath(__file__))
 try:
    import xformers
    import xformers.ops
    XFORMERS_IS_AVAILABLE = True
 except:
    XFORMERS_IS_AVAILABLE = False
 def dummy_build_vision_tower(*args, **kwargs):
    # Monkey patch the CLIP class before you create an instance.
    return None
-open_clip.model._build_vision_tower = dummy_build_vision_tower
+
@contextmanager
 def patch_build_vision_tower():
    original_build_vision_tower = open_clip.model._build_vision_tower
    open_clip.model._build_vision_tower = dummy_build_vision_tower
    try:
        yield
    finally:
        open_clip.model._build_vision_tower = original_build_vision_tower
 def build_text_model_from_openai_state_dict(
        state_dict: dict,
        cast_dtype=torch.float16,
-):
+    ):
    embed_dim = state_dict["text_projection"].shape[1]
    context_length = state_dict["positional_embedding"].shape[0]
@@ -56,13 +55,15 @@ def build_text_model_from_openai_state_dict(
        heads=transformer_heads,
        layers=transformer_layers,
    )
-    model = open_clip.CLIP(
+
-        embed_dim,
+    with patch_build_vision_tower():
-        vision_cfg=vision_cfg,
+        model = open_clip.CLIP(
-        text_cfg=text_cfg,
+            embed_dim,
-        quick_gelu=True,  # OpenAI models were trained with QuickGELU
+            vision_cfg=vision_cfg,
-        cast_dtype=cast_dtype,
+            text_cfg=text_cfg,
-    )
+            quick_gelu=True,
            cast_dtype=cast_dtype,
        )
    model.load_state_dict(state_dict, strict=False)
    model = model.eval()
@@ -84,13 +85,13 @@ class SUPIR_Upscale:
            "scale_by": ("FLOAT", {"default": 1.0, "min": 0.01, "max": 20.0, "step": 0.01}),
            "steps": ("INT", {"default": 45, "min": 3, "max": 4096, "step": 1}),
            "restoration_scale": ("FLOAT", {"default": -1.0, "min": -1.0, "max": 6.0, "step": 1.0}),
-            "cfg_scale": ("FLOAT", {"default": 4.0, "min": 0, "max": 20, "step": 0.01}),
+            "cfg_scale": ("FLOAT", {"default": 4.0, "min": 0, "max": 100, "step": 0.01}),
            "a_prompt": ("STRING", {"multiline": True, "default": "high quality, detailed", }),
            "n_prompt": ("STRING", {"multiline": True, "default": "bad quality, blurry, messy", }),
            "s_churn": ("INT", {"default": 5, "min": 0, "max": 40, "step": 1}),
            "s_noise": ("FLOAT", {"default": 1.003, "min": 1.0, "max": 1.1, "step": 0.001}),
            "control_scale": ("FLOAT", {"default": 1.0, "min": 0, "max": 10.0, "step": 0.05}),
-            "cfg_scale_start": ("FLOAT", {"default": 4.0, "min": 0.0, "max": 9.0, "step": 0.05}),
+            "cfg_scale_start": ("FLOAT", {"default": 4.0, "min": 0.0, "max": 100.0, "step": 0.05}),
            "control_scale_start": ("FLOAT", {"default": 0.0, "min": 0, "max": 1.0, "step": 0.05}),
            "color_fix_type": (
                [
@@ -128,6 +129,15 @@ class SUPIR_Upscale:
                "use_tiled_sampling": ("BOOLEAN", {"default": False}),
                "sampler_tile_size": ("INT", {"default": 1024, "min": 64, "max": 4096, "step": 32}),
                "sampler_tile_stride": ("INT", {"default": 512, "min": 32, "max": 2048, "step": 32}),
                "fp8_unet": ("BOOLEAN", {"default": False}),
                "fp8_vae": ("BOOLEAN", {"default": False}),
                "sampler": (
                    [
                        'RestoreDPMPP2MSampler',
                        'RestoreEDMSampler',
                    ], {
                        "default": 'RestoreEDMSampler'
                    }),
            }
        }
@@ -141,7 +151,7 @@ class SUPIR_Upscale:
                encoder_tile_size_pixels, decoder_tile_size_latent,
                control_scale, cfg_scale_start, control_scale_start, restoration_scale, keep_model_loaded,
                a_prompt, n_prompt, sdxl_model, supir_model, use_tiled_vae, use_tiled_sampling=False, sampler_tile_size=128, sampler_tile_stride=64, captions="", diffusion_dtype="auto",
-                encoder_dtype="auto", batch_size=1):
+                encoder_dtype="auto", batch_size=1, fp8_unet=False, fp8_vae=False, sampler="RestoreEDMSampler"):
        device = mm.get_torch_device()
        mm.unload_all_models()
@@ -160,18 +170,21 @@ class SUPIR_Upscale:
            'use_tiled_vae': use_tiled_vae,
            'supir_model': supir_model,
            'use_tiled_sampling': use_tiled_sampling,
            'fp8_unet': fp8_unet,
            'fp8_vae': fp8_vae,
            'sampler': sampler
        }
        if diffusion_dtype == 'auto':
            try:
                if mm.should_use_fp16():
                    print("Diffusion using fp16")
                    dtype = torch.float16
                    model_dtype = 'fp16'
                if mm.should_use_bf16():
                    print("Diffusion using bf16")
                    dtype = torch.bfloat16
                    model_dtype = 'bf16'
                elif mm.should_use_fp16():
                    print("Diffusion using using fp16")
                    dtype = torch.float16
                    model_dtype = 'fp16'
                else:
                    print("Diffusion using using fp32")
                    dtype = torch.float32
@@ -207,19 +220,21 @@ class SUPIR_Upscale:
                config = OmegaConf.load(config_path_tiled)
                config.model.params.sampler_config.params.tile_size = sampler_tile_size // 8
                config.model.params.sampler_config.params.tile_stride = sampler_tile_stride // 8
                config.model.params.sampler_config.target = f".sgm.modules.diffusionmodules.sampling.Tiled{sampler}"
                print("Using tiled sampling")
            else:
                config = OmegaConf.load(config_path)
                config.model.params.sampler_config.target = f".sgm.modules.diffusionmodules.sampling.{sampler}"
                print("Using non-tiled sampling")
-            if XFORMERS_IS_AVAILABLE:
+            if mm.XFORMERS_IS_AVAILABLE:
                config.model.params.control_stage_config.params.spatial_transformer_attn_type = "softmax-xformers"
                config.model.params.network_config.params.spatial_transformer_attn_type = "softmax-xformers"
                config.model.params.first_stage_config.params.ddconfig.attn_type = "vanilla-xformers" 
            config.model.params.ae_dtype = vae_dtype
            config.model.params.diffusion_dtype = model_dtype
-
+            
            self.model = instantiate_from_config(config.model).cpu()
            try:
@@ -267,35 +282,30 @@ class SUPIR_Upscale:
                clip_g = build_text_model_from_openai_state_dict(sd, cast_dtype=dtype)
                self.model.conditioner.embedders[1].model = clip_g
            except:
                raise Exception("Failed to load second clip model from SDXL checkpoint")
            del sd, clip_g
            mm.soft_empty_cache()
-            try:
+            self.model.to(dtype)
-                self.model.to(dtype)
+
-                self.model.to(device)
+            #only unets and/or vae to fp8 
-            except Exception as e:
+            if fp8_unet:
-                print("Failed to move model to device")
+                self.model.model.to(torch.float8_e4m3fn)
-                print(e)
+            if fp8_vae:
-                import gc
+                self.model.first_stage_model.to(torch.float8_e4m3fn)
                # unload everything and give up
                self.model = None
                del self.model
                gc.collect()
                mm.soft_empty_cache()
            if use_tiled_vae:
                self.model.init_tile_vae(encoder_tile_size=encoder_tile_size_pixels, decoder_tile_size=decoder_tile_size_latent)
-        image, = ImageScaleBy.upscale(self, image, resize_method, scale_by)
+        upscaled_image, = ImageScaleBy.upscale(self, image, resize_method, scale_by)
-        B, H, W, C = image.shape
+        B, H, W, C = upscaled_image.shape
-        new_height = H // 64 * 64
+        new_height = H if H % 64 == 0 else ((H // 64) + 1) * 64
-        new_width = W // 64 * 64
+        new_width = W if W % 64 == 0 else ((W // 64) + 1) * 64
-        image = image.permute(0, 3, 1, 2).contiguous()
+        upscaled_image = upscaled_image.permute(0, 3, 1, 2)
-        resized_image = F.interpolate(image, size=(new_height, new_width), mode='bicubic', align_corners=False)
+        resized_image = F.interpolate(upscaled_image, size=(new_height, new_width), mode='bicubic', align_corners=False)
        resized_image = resized_image.to(device)
        captions_list = []
        captions_list.append(captions)
        print("captions: ", captions_list)
@@ -329,7 +339,8 @@ class SUPIR_Upscale:
                self.model = None
                mm.soft_empty_cache()
                print("It's likely that too large of an image or batch_size for SUPIR was used,"
-                      " and it has devoured all of the memory it had reserved, you may need to restart ComfyUI")
+                      " and it has devoured all of the memory it had reserved, you may need to restart ComfyUI. Make sure you are using tiled_vae, "
                      " you can also try using fp8 for reduced memory usage if your system supports it.")
                raise e
            out.append(samples.squeeze(0).cpu())
@@ -345,7 +356,7 @@ class SUPIR_Upscale:
        else:
            out_stacked = torch.stack(out, dim=0).cpu().to(torch.float32).permute(0, 2, 3, 1)
-        final_image, = ImageScale.upscale(self, out_stacked, "lanczos", W, H, crop="disabled")
+        final_image, = ImageScale.upscale(self, out_stacked, resize_method, W, H, crop="disabled")
        return (final_image,)
@@ -354,4 +365,4 @@ NODE_CLASS_MAPPINGS = {
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
    "SUPIR_Upscale": "SUPIR_Upscale"
-}
+}
@@ -0,0 +1,14 @@
 [project]
 name = "comfyui-supir"
 description = "Wrapper nodes to use SUPIR upscaling process in ComfyUI"
 version = "1.0.4"
 license = { file = "LICENSE" }
 dependencies = ["transformers>=4.28.1", "fsspec>=2023.4.0", "kornia>=0.6.9", "open-clip-torch>=2.24.0", "Pillow>=9.4.0", "pytorch-lightning>=2.2.1", "omegaconf", "accelerate"]
 [project.urls]
 Repository = "https://github.com/kijai/ComfyUI-SUPIR"
 [tool.comfy]
 PublisherId = "kijai"
 DisplayName = "ComfyUI-SUPIR"
 Icon = ""
@@ -1,8 +1,6 @@
 transformers>=4.28.1
-openai-clip>=1.0.1
+open-clip-torch>=2.24.0
 fsspec>=2023.4.0
 kornia>=0.6.9
 open-clip-torch>=2.17.1
 Pillow>=9.4.0
-pytorch-lightning==2.1.2
+pytorch-lightning>=2.5.5
-omegaconf
+omegaconf
 accelerate
@@ -14,9 +14,8 @@ from ..modules.distributions.distributions import DiagonalGaussianDistribution
 from ..modules.ema import LitEma
 from ..util import default, get_obj_from_str, instantiate_from_config
-class Conv2d(torch.nn.Conv2d):
+import comfy.ops
-    def reset_parameters(self):
+ops = comfy.ops.manual_cast
        return None
 class AbstractAutoencoder(pl.LightningModule):
    """
@@ -297,8 +296,8 @@ class AutoencoderKL(AutoencodingEngine):
        assert ddconfig["double_z"]
        self.encoder = Encoder(**ddconfig)
        self.decoder = Decoder(**ddconfig)
-        self.quant_conv = Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1)
+        self.quant_conv = ops.Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1)
-        self.post_quant_conv = Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.post_quant_conv = ops.Conv2d(embed_dim, ddconfig["z_channels"], 1)
        self.embed_dim = embed_dim
        if ckpt_path is not None:
@@ -10,13 +10,8 @@ from einops import rearrange, repeat
 from packaging import version
 from torch import nn
-class Conv2d(torch.nn.Conv2d):
+import comfy.ops
-    def reset_parameters(self):
+ops = comfy.ops.manual_cast
        return None
 class Linear(torch.nn.Linear):
    def reset_parameters(self):
        return None
 if version.parse(torch.__version__) >= version.parse("2.0.0"):
    SDP_IS_AVAILABLE = True
@@ -92,7 +87,7 @@ def init_(tensor):
 class GEGLU(nn.Module):
    def __init__(self, dim_in, dim_out):
        super().__init__()
-        self.proj = Linear(dim_in, dim_out * 2)
+        self.proj = ops.Linear(dim_in, dim_out * 2)
    def forward(self, x):
        x, gate = self.proj(x).chunk(2, dim=-1)
@@ -105,13 +100,13 @@ class FeedForward(nn.Module):
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
        project_in = (
-            nn.Sequential(Linear(dim, inner_dim), nn.GELU())
+            nn.Sequential(ops.Linear(dim, inner_dim), nn.GELU())
            if not glu
            else GEGLU(dim, inner_dim)
        )
        self.net = nn.Sequential(
-            project_in, nn.Dropout(dropout), Linear(inner_dim, dim_out)
+            project_in, nn.Dropout(dropout), ops.Linear(inner_dim, dim_out)
        )
    def forward(self, x):
@@ -128,7 +123,7 @@ def zero_module(module):
 def Normalize(in_channels):
-    return torch.nn.GroupNorm(
+    return ops.GroupNorm(
        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
    )
@@ -138,8 +133,8 @@ class LinearAttention(nn.Module):
        super().__init__()
        self.heads = heads
        hidden_dim = dim_head * heads
-        self.to_qkv = Conv2d(dim, hidden_dim * 3, 1, bias=False)
+        self.to_qkv = ops.Conv2d(dim, hidden_dim * 3, 1, bias=False)
-        self.to_out = Conv2d(hidden_dim, dim, 1)
+        self.to_out = ops.Conv2d(hidden_dim, dim, 1)
    def forward(self, x):
        b, c, h, w = x.shape
@@ -162,16 +157,16 @@ class SpatialSelfAttention(nn.Module):
        self.in_channels = in_channels
        self.norm = Normalize(in_channels)
-        self.q = Conv2d(
+        self.q = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.k = Conv2d(
+        self.k = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.v = Conv2d(
+        self.v = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.proj_out = Conv2d(
+        self.proj_out = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
@@ -218,12 +213,12 @@ class CrossAttention(nn.Module):
        self.scale = dim_head**-0.5
        self.heads = heads
-        self.to_q = Linear(query_dim, inner_dim, bias=False)
+        self.to_q = ops.Linear(query_dim, inner_dim, bias=False)
-        self.to_k = Linear(context_dim, inner_dim, bias=False)
+        self.to_k = ops.Linear(context_dim, inner_dim, bias=False)
-        self.to_v = Linear(context_dim, inner_dim, bias=False)
+        self.to_v = ops.Linear(context_dim, inner_dim, bias=False)
        self.to_out = nn.Sequential(
-            Linear(inner_dim, query_dim), nn.Dropout(dropout)
+            ops.Linear(inner_dim, query_dim), nn.Dropout(dropout)
        )
        self.backend = backend
@@ -309,12 +304,12 @@ class MemoryEfficientCrossAttention(nn.Module):
        self.heads = heads
        self.dim_head = dim_head
-        self.to_q = Linear(query_dim, inner_dim, bias=False)
+        self.to_q = ops.Linear(query_dim, inner_dim, bias=False)
-        self.to_k = Linear(context_dim, inner_dim, bias=False)
+        self.to_k = ops.Linear(context_dim, inner_dim, bias=False)
-        self.to_v = Linear(context_dim, inner_dim, bias=False)
+        self.to_v = ops.Linear(context_dim, inner_dim, bias=False)
        self.to_out = nn.Sequential(
-            Linear(inner_dim, query_dim), nn.Dropout(dropout)
+            ops.Linear(inner_dim, query_dim), nn.Dropout(dropout)
        )
        self.attention_op: Optional[Any] = None
@@ -442,9 +437,9 @@ class BasicTransformerBlock(nn.Module):
            dropout=dropout,
            backend=sdp_backend,
        )  # is self-attn if context is none
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = ops.LayerNorm(dim)
-        self.norm2 = nn.LayerNorm(dim)
+        self.norm2 = ops.LayerNorm(dim)
-        self.norm3 = nn.LayerNorm(dim)
+        self.norm3 = ops.LayerNorm(dim)
        self.checkpoint = checkpoint
        #if self.checkpoint:
            #print(f"{self.__class__.__name__} is using checkpointing")
@@ -523,8 +518,8 @@ class BasicTransformerSingleLayerBlock(nn.Module):
            context_dim=context_dim,
        )
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = ops.LayerNorm(dim)
-        self.norm2 = nn.LayerNorm(dim)
+        self.norm2 = ops.LayerNorm(dim)
        self.checkpoint = checkpoint
    def forward(self, x, context=None):
@@ -564,9 +559,9 @@ class SpatialTransformer(nn.Module):
        sdp_backend=None,
    ):
        super().__init__()
-        print(
+        # print(
-            f"constructing {self.__class__.__name__} of depth {depth} w/ {in_channels} channels and {n_heads} heads"
+        #     f"constructing {self.__class__.__name__} of depth {depth} w/ {in_channels} channels and {n_heads} heads"
-        )
+        # )
        from omegaconf import ListConfig
        if exists(context_dim) and not isinstance(context_dim, (list, ListConfig)):
@@ -588,11 +583,11 @@ class SpatialTransformer(nn.Module):
        inner_dim = n_heads * d_head
        self.norm = Normalize(in_channels)
        if not use_linear:
-            self.proj_in = Conv2d(
+            self.proj_in = ops.Conv2d(
                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
            )
        else:
-            self.proj_in = Linear(in_channels, inner_dim)
+            self.proj_in = ops.Linear(in_channels, inner_dim)
        self.transformer_blocks = nn.ModuleList(
            [
@@ -612,11 +607,11 @@ class SpatialTransformer(nn.Module):
        )
        if not use_linear:
            self.proj_out = zero_module(
-                Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+                ops.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
            )
        else:
            # self.proj_out = zero_module(Linear(in_channels, inner_dim))
-            self.proj_out = zero_module(Linear(inner_dim, in_channels))
+            self.proj_out = zero_module(ops.Linear(inner_dim, in_channels))
        self.use_linear = use_linear
    def forward(self, x, context=None):
@@ -8,6 +8,8 @@ from torchvision import models
 from ..util import get_ckpt_path
 import comfy.ops
 ops = comfy.ops.manual_cast
 class LPIPS(nn.Module):
    # Learned perceptual metric
@@ -91,7 +93,7 @@ class NetLinLayer(nn.Module):
            else []
        )
        layers += [
-            nn.Conv2d(chn_in, chn_out, 1, stride=1, padding=0, bias=False),
+            ops.Conv2d(chn_in, chn_out, 1, stride=1, padding=0, bias=False),
        ]
        self.model = nn.Sequential(*layers)
@@ -3,7 +3,8 @@ import functools
 import torch.nn as nn
 from ..util import ActNorm
-
+import comfy.ops
 ops = comfy.ops.manual_cast
 def weights_init(m):
    classname = m.__class__.__name__
@@ -42,7 +43,7 @@ class NLayerDiscriminator(nn.Module):
        kw = 4
        padw = 1
        sequence = [
-            nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw),
+            ops.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw),
            nn.LeakyReLU(0.2, True),
        ]
        nf_mult = 1
@@ -51,7 +52,7 @@ class NLayerDiscriminator(nn.Module):
            nf_mult_prev = nf_mult
            nf_mult = min(2**n, 8)
            sequence += [
-                nn.Conv2d(
+                ops.Conv2d(
                    ndf * nf_mult_prev,
                    ndf * nf_mult,
                    kernel_size=kw,
@@ -66,7 +67,7 @@ class NLayerDiscriminator(nn.Module):
        nf_mult_prev = nf_mult
        nf_mult = min(2**n_layers, 8)
        sequence += [
-            nn.Conv2d(
+            ops.Conv2d(
                ndf * nf_mult_prev,
                ndf * nf_mult,
                kernel_size=kw,
@@ -79,7 +80,7 @@ class NLayerDiscriminator(nn.Module):
        ]
        sequence += [
-            nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)
+            ops.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)
        ]  # output 1 channel prediction map
        self.main = nn.Sequential(*sequence)
@@ -19,13 +19,8 @@ except:
 from ...modules.attention import LinearAttention, MemoryEfficientCrossAttention
-class Conv2d(torch.nn.Conv2d):
+import comfy.ops
-    def reset_parameters(self):
+ops = comfy.ops.manual_cast
        return None
 class Linear(torch.nn.Linear):
    def reset_parameters(self):
        return None
 def get_timestep_embedding(timesteps, embedding_dim):
    """
@@ -54,7 +49,7 @@ def nonlinearity(x):
 def Normalize(in_channels, num_groups=32):
-    return torch.nn.GroupNorm(
+    return ops.GroupNorm(
        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
    )
@@ -64,7 +59,7 @@ class Upsample(nn.Module):
        super().__init__()
        self.with_conv = with_conv
        if self.with_conv:
-            self.conv = Conv2d(
+            self.conv = ops.Conv2d(
                in_channels, in_channels, kernel_size=3, stride=1, padding=1
            )
@@ -91,7 +86,7 @@ class Downsample(nn.Module):
        self.with_conv = with_conv
        if self.with_conv:
            # no asymmetric padding in torch conv, must do it ourselves
-            self.conv = Conv2d(
+            self.conv = ops.Conv2d(
                in_channels, in_channels, kernel_size=3, stride=2, padding=0
            )
@@ -122,23 +117,23 @@ class ResnetBlock(nn.Module):
        self.use_conv_shortcut = conv_shortcut
        self.norm1 = Normalize(in_channels)
-        self.conv1 = Conv2d(
+        self.conv1 = ops.Conv2d(
            in_channels, out_channels, kernel_size=3, stride=1, padding=1
        )
        if temb_channels > 0:
-            self.temb_proj = Linear(temb_channels, out_channels)
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
        self.norm2 = Normalize(out_channels)
        self.dropout = torch.nn.Dropout(dropout)
-        self.conv2 = Conv2d(
+        self.conv2 = ops.Conv2d(
            out_channels, out_channels, kernel_size=3, stride=1, padding=1
        )
        if self.in_channels != self.out_channels:
            if self.use_conv_shortcut:
-                self.conv_shortcut = Conv2d(
+                self.conv_shortcut = ops.Conv2d(
                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
                )
            else:
-                self.nin_shortcut = Conv2d(
+                self.nin_shortcut = ops.Conv2d(
                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
                )
@@ -178,16 +173,16 @@ class AttnBlock(nn.Module):
        self.in_channels = in_channels
        self.norm = Normalize(in_channels)
-        self.q = Conv2d(
+        self.q = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.k = Conv2d(
+        self.k = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.v = Conv2d(
+        self.v = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.proj_out = Conv2d(
+        self.proj_out = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
@@ -228,16 +223,16 @@ class MemoryEfficientAttnBlock(nn.Module):
        self.in_channels = in_channels
        self.norm = Normalize(in_channels)
-        self.q = Conv2d(
+        self.q = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.k = Conv2d(
+        self.k = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.v = Conv2d(
+        self.v = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
-        self.proj_out = Conv2d(
+        self.proj_out = ops.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )
        self.attention_op: Optional[Any] = None
@@ -354,13 +349,13 @@ class Model(nn.Module):
            self.temb = nn.Module()
            self.temb.dense = nn.ModuleList(
                [
-                    Linear(self.ch, self.temb_ch),
+                    ops.Linear(self.ch, self.temb_ch),
-                    Linear(self.temb_ch, self.temb_ch),
+                    ops.Linear(self.temb_ch, self.temb_ch),
                ]
            )
        # downsampling
-        self.conv_in = Conv2d(
+        self.conv_in = ops.Conv2d(
            in_channels, self.ch, kernel_size=3, stride=1, padding=1
        )
@@ -439,7 +434,7 @@ class Model(nn.Module):
        # end
        self.norm_out = Normalize(block_in)
-        self.conv_out = Conv2d(
+        self.conv_out = ops.Conv2d(
            block_in, out_ch, kernel_size=3, stride=1, padding=1
        )
@@ -526,7 +521,7 @@ class Encoder(nn.Module):
        self.in_channels = in_channels
        # downsampling
-        self.conv_in = Conv2d(
+        self.conv_in = ops.Conv2d(
            in_channels, self.ch, kernel_size=3, stride=1, padding=1
        )
@@ -577,7 +572,7 @@ class Encoder(nn.Module):
        # end
        self.norm_out = Normalize(block_in)
-        self.conv_out = Conv2d(
+        self.conv_out = ops.Conv2d(
            block_in,
            2 * z_channels if double_z else z_channels,
            kernel_size=3,
@@ -660,7 +655,7 @@ class Decoder(nn.Module):
        make_resblock_cls = self._make_resblock()
        make_conv_cls = self._make_conv()
        # z to block_in
-        self.conv_in = Conv2d(
+        self.conv_in = ops.Conv2d(
            z_channels, block_in, kernel_size=3, stride=1, padding=1
        )
@@ -719,7 +714,7 @@ class Decoder(nn.Module):
        return ResnetBlock
    def _make_conv(self) -> Callable:
-        return Conv2d
+        return ops.Conv2d
    def get_last_layer(self, **kwargs):
        return self.conv_out.weight
@@ -186,13 +186,13 @@ class Downsample(nn.Module):
        self.dims = dims
        stride = 2 if dims != 3 else ((1, 2, 2) if not third_down else (2, 2, 2))
        if use_conv:
-            print(f"Building a Downsample layer with {dims} dims.")
+            # print(f"Building a Downsample layer with {dims} dims.")
-            print(
+            # print(
-                f"  --> settings are: \n in-chn: {self.channels}, out-chn: {self.out_channels}, "
+            #     f"  --> settings are: \n in-chn: {self.channels}, out-chn: {self.out_channels}, "
-                f"kernel-size: 3, stride: {stride}, padding: {padding}"
+            #     f"kernel-size: 3, stride: {stride}, padding: {padding}"
-            )
+            # )
-            if dims == 3:
+            # if dims == 3:
-                print(f"  --> Downsampling third axis (time): {third_down}")
+            #     print(f"  --> Downsampling third axis (time): {third_down}")
            self.op = conv_nd(
                dims,
                self.channels,
@@ -8,7 +8,8 @@ from typing import Dict, Union
 import torch
 from omegaconf import ListConfig, OmegaConf
 from tqdm import tqdm
-
+from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras
 from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras
 from ...modules.diffusionmodules.sampling_utils import (
    get_ancestral_step,
    linear_multistep_coeff,
@@ -17,6 +18,7 @@ from ...modules.diffusionmodules.sampling_utils import (
    to_sigma,
 )
 from ...util import append_dims, default, instantiate_from_config
 import copy
 DEFAULT_GUIDER = {"target": ".sgm.modules.diffusionmodules.guiders.IdentityGuider"}
@@ -461,18 +463,20 @@ class TiledRestoreEDMSampler(RestoreEDMSampler):
    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, x_center=None, control_scale=1.0,
                 use_linear_control_scale=False, control_scale_start=0.0):
-        use_local_prompt = isinstance(cond, list)
+        cond_copy = copy.deepcopy(cond)
        uc_copy = copy.deepcopy(uc)
        use_local_prompt = isinstance(cond_copy, list)
        b, _, h, w = x.shape
        latent_tiles_iterator = _sliding_windows(h, w, self.tile_size, self.tile_stride)
        tile_weights = self.tile_weights.repeat(b, 1, 1, 1)
        if not use_local_prompt:
-            LQ_latent = cond['control']
+            LQ_latent = cond_copy['control']
        else:
-            assert len(cond) == len(latent_tiles_iterator), "Number of local prompts should be equal to number of tiles"
+            assert len(cond_copy) == len(latent_tiles_iterator), "Number of local prompts should be equal to number of tiles"
-            LQ_latent = cond[0]['control']
+            LQ_latent = cond_copy[0]['control']
        clean_LQ_latent = x_center
-        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
+        x, s_in, sigmas, num_sigmas, cond_copy, uc_copy = self.prepare_sampling_loop(
-            x, cond, uc, num_steps
+            x, cond_copy, uc_copy, num_steps
        )
        pbar_comfy = comfy.utils.ProgressBar(num_sigmas)
        for _idx, i in enumerate(self.get_sigma_gen(num_sigmas)):
@@ -489,18 +493,18 @@ class TiledRestoreEDMSampler(RestoreEDMSampler):
                _eps_noise = eps_noise[:, :, hi:hi_end, wi:wi_end]
                x_center_tile = clean_LQ_latent[:, :, hi:hi_end, wi:wi_end]
                if use_local_prompt:
-                    _cond = cond[j]
+                    _cond = cond_copy[j]
                else:
-                    _cond = cond
+                    _cond = cond_copy
                _cond['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
-                uc['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
+                uc_copy['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
                _x = self.sampler_step(
                    s_in * sigmas[i],
                    s_in * sigmas[i + 1],
                    denoiser,
                    x_tile,
                    _cond,
-                    uc,
+                    uc_copy,
                    gamma,
                    x_center_tile,
                    eps_noise=_eps_noise,
@@ -555,3 +559,193 @@ def _sliding_windows(h: int, w: int, tile_size: int, tile_stride: int):
            coords.append((hi, hi + tile_size, wi, wi + tile_size))
    return coords
 class RestoreDPMPP2MSampler(DPMPP2MSampler):
    def __init__(self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, restore_cfg=4.0,
            restore_cfg_s_tmin=0.05, eta=1., *args, **kwargs):
        self.s_noise = s_noise
        self.eta = eta
        self.restore_cfg = restore_cfg
        self.restore_cfg_s_tmin = restore_cfg_s_tmin
        self.sigma_max = 14.6146
        super().__init__(*args, **kwargs)
    def denoise(self, x, denoiser, sigma, cond, uc, control_scale=1.0):
        denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc), control_scale)
        denoised = self.guider(denoised, sigma)
        return denoised
    def get_mult(self, h, r, t, t_next, previous_sigma):
        eta_h = self.eta * h
        mult1 = to_sigma(t_next) / to_sigma(t) * (-eta_h).exp()
        mult2 = (-h -eta_h).expm1()
        if previous_sigma is not None:
            mult3 = 1 + 1 / (2 * r)
            mult4 = 1 / (2 * r)
            return mult1, mult2, mult3, mult4
        else:
            return mult1, mult2
    def sampler_step(
        self,
        old_denoised,
        previous_sigma,
        sigma,
        next_sigma,
        denoiser,
        x,
        cond,
        uc=None,
        eps_noise=None,
        x_center=None,
        control_scale=1.0,
        use_linear_control_scale=False,
        control_scale_start=0.0
    ):
        if use_linear_control_scale:
            control_scale = (sigma[0].item() / self.sigma_max) * (control_scale_start - control_scale) + control_scale
        denoised = self.denoise(x, denoiser, sigma, cond, uc, control_scale=control_scale)
        if (next_sigma[0] > self.restore_cfg_s_tmin) and (self.restore_cfg > 0):
            d_center = (denoised - x_center)
            denoised = denoised - d_center * ((sigma.view(-1, 1, 1, 1) / self.sigma_max) ** self.restore_cfg)
        h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma)
        eta_h = self.eta * h
        mult = [
            append_dims(mult, x.ndim)
            for mult in self.get_mult(h, r, t, t_next, previous_sigma)
        ]
        x_standard = mult[0] * x - mult[1] * denoised
        if old_denoised is None or torch.sum(next_sigma) < 1e-14:
            # Save a network evaluation if all noise levels are 0 or on the first step
            return x_standard, denoised
        else:
            denoised_d = mult[2] * denoised - mult[3] * old_denoised
            x_advanced = mult[0] * x - mult[1] * denoised_d
            # apply correction if noise level is not 0 and not first step
            x = torch.where(
                append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard
            )
            if self.eta:
                x = x + eps_noise * next_sigma * (-2 * eta_h).expm1().neg().sqrt() * self.s_noise
        return x, denoised
    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, x_center=None, control_scale=1.0, 
                 use_linear_control_scale=False, control_scale_start=0.0, **kwargs):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )
        sigmas_min, sigmas_max = sigmas[-2].cpu(), sigmas[0].cpu()
        sigmas_new = get_sigmas_karras(self.num_steps, sigmas_min, sigmas_max, device=x.device)
        sigmas = sigmas_new
        noise_sampler = BrownianTreeNoiseSampler(x, sigmas_min, sigmas_max)
        old_denoised = None
        pbar_comfy = comfy.utils.ProgressBar(num_sigmas)
        for i in self.get_sigma_gen(num_sigmas):
            if i > 0 and torch.sum(s_in * sigmas[i + 1]) > 1e-14:
                eps_noise = noise_sampler(s_in * sigmas[i], s_in * sigmas[i + 1])
            else:
                eps_noise = None
            x, old_denoised = self.sampler_step(
                old_denoised,
                None if i == 0 else s_in * sigmas[i - 1],
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc=uc,
                eps_noise=eps_noise,
                control_scale=control_scale,
                x_center=x_center,
                use_linear_control_scale=use_linear_control_scale,
                control_scale_start=control_scale_start,
            )
            pbar_comfy.update(1)
        return x
 class TiledRestoreDPMPP2MSampler(RestoreDPMPP2MSampler):
    def __init__(self, tile_size=128, tile_stride=64, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.tile_size = tile_size
        self.tile_stride = tile_stride
        self.tile_weights = gaussian_weights(self.tile_size, self.tile_size, 1)
    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, control_scale=1.0, **kwargs):
        use_local_prompt = isinstance(cond, list)
        b, _, h, w = x.shape
        latent_tiles_iterator = _sliding_windows(h, w, self.tile_size, self.tile_stride)
        print(f"Image divided into {len(latent_tiles_iterator)} tiles")
        print("Conds received: ", len(cond))
        cond_copy = copy.deepcopy(cond)
        uc_copy = copy.deepcopy(uc)
        tile_weights = self.tile_weights.repeat(b, 1, 1, 1)
        if not use_local_prompt:
            LQ_latent = cond['control']
        else:
            assert len(cond_copy) == len(latent_tiles_iterator), "Number of local prompts should be equal to number of tiles"
            LQ_latent = cond_copy[0]['control']
            print("LQ_latent shape: ",LQ_latent.shape)
        x, s_in, sigmas, num_sigmas, cond_copy, uc_copy = self.prepare_sampling_loop(
            x, cond_copy, uc_copy, num_steps
        )
        sigmas_min, sigmas_max = sigmas[-2].cpu(), sigmas[0].cpu()
        sigmas_new = get_sigmas_karras(self.num_steps, sigmas_min, sigmas_max, device=x.device)
        sigmas = sigmas_new
        noise_sampler = BrownianTreeNoiseSampler(x, sigmas_min, sigmas_max)
        old_denoised = None
        pbar_comfy = comfy.utils.ProgressBar(num_sigmas)
        for _idx, i in enumerate(self.get_sigma_gen(num_sigmas)):
            if i > 0 and torch.sum(s_in * sigmas[i + 1]) > 1e-14:
                eps_noise = noise_sampler(s_in * sigmas[i], s_in * sigmas[i + 1])
            else:
                eps_noise = torch.zeros_like(x)
            x_next = torch.zeros_like(x)
            old_denoised_next = torch.zeros_like(x)
            count = torch.zeros_like(x)
            for j, (hi, hi_end, wi, wi_end) in enumerate(latent_tiles_iterator):
                x_tile = x[:, :, hi:hi_end, wi:wi_end]
                _eps_noise = eps_noise[:, :, hi:hi_end, wi:wi_end]
                if old_denoised is not None:
                    old_denoised_tile = old_denoised[:, :, hi:hi_end, wi:wi_end]
                else:
                    old_denoised_tile = None
                if use_local_prompt:
                    _cond = cond_copy[j]
                else:
                    _cond = cond_copy
                _cond['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
                uc_copy['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
                _x, _old_denoised = self.sampler_step(
                    old_denoised_tile,
                    None if i == 0 else s_in * sigmas[i - 1],
                    s_in * sigmas[i],
                    s_in * sigmas[i + 1],
                    denoiser,
                    x_tile,
                    _cond,
                    uc=uc_copy,
                    eps_noise=_eps_noise,
                    control_scale=control_scale,
                )
                x_next[:, :, hi:hi_end, wi:wi_end] += _x * tile_weights
                old_denoised_next[:, :, hi:hi_end, wi:wi_end] += _old_denoised * tile_weights
                count[:, :, hi:hi_end, wi:wi_end] += tile_weights
            old_denoised_next /= count
            x_next /= count
            x = x_next
            old_denoised = old_denoised_next
            pbar_comfy.update(1)
        return x
@@ -19,6 +19,9 @@ import comfy.model_management
 device = comfy.model_management.get_torch_device()
 from contextlib import nullcontext
 import comfy.ops
 ops = comfy.ops.manual_cast
 def make_beta_schedule(
    schedule,
    n_timestep,
@@ -278,25 +281,17 @@ class GroupNorm32(nn.GroupNorm):
    def forward(self, x):
        # return super().forward(x.float()).type(x.dtype)
        return super().forward(x)
 class Conv2d(torch.nn.Conv2d):
    def reset_parameters(self):
        return None
 class Linear(torch.nn.Linear):
    def reset_parameters(self):
        return None
 def conv_nd(dims, *args, **kwargs):
    """
    Create a 1D, 2D, or 3D convolution module.
    """
    if dims == 1:
-        return nn.Conv1d(*args, **kwargs)
+        return ops.Conv1d(*args, **kwargs)
    elif dims == 2:
-        return Conv2d(*args, **kwargs)
+        return ops.Conv2d(*args, **kwargs)
    elif dims == 3:
-        return nn.Conv3d(*args, **kwargs)
+        return ops.Conv3d(*args, **kwargs)
    raise ValueError(f"unsupported dimensions: {dims}")
@@ -304,7 +299,7 @@ def linear(*args, **kwargs):
    """
    Create a linear module.
    """
-    return Linear(*args, **kwargs)
+    return ops.Linear(*args, **kwargs)
 def avg_pool_nd(dims, *args, **kwargs):
@@ -2,7 +2,7 @@ from contextlib import nullcontext
 from functools import partial
 from typing import Dict, List, Optional, Tuple, Union
-import kornia
+#import kornia
 import numpy as np
 import open_clip
 import torch
@@ -36,6 +36,9 @@ from ....CKPT_PTH import SDXL_CLIP1_PATH, SDXL_CLIP2_CKPT_PTH
 import comfy.model_management
 device = comfy.model_management.get_torch_device()
 import comfy.ops
 ops = comfy.ops.manual_cast
 class AbstractEmbModel(nn.Module):
    def __init__(self):
        super().__init__()
@@ -99,10 +102,10 @@ class GeneralConditioner(nn.Module):
                for param in embedder.parameters():
                    param.requires_grad = False
                embedder.eval()
-            print(
+            # print(
-                f"Initialized embedder #{n}: {embedder.__class__.__name__} "
+            #     f"Initialized embedder #{n}: {embedder.__class__.__name__} "
-                f"with {count_params(embedder, False)} params. Trainable: {embedder.is_trainable}"
+            #     f"with {count_params(embedder, False)} params. Trainable: {embedder.is_trainable}"
-            )
+            # )
            if "input_key" in embconfig:
                embedder.input_key = embconfig["input_key"]
@@ -577,7 +580,10 @@ class FrozenOpenCLIPEmbedder2(AbstractEmbModel):
        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
        x = x + self.model.positional_embedding
        x = x.permute(1, 0, 2)  # NLD -> LND
-        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
+        try:
            x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
        except:
            x = self.text_transformer_forward_batch_first(x, attn_mask=self.model.attn_mask)
        if self.legacy:
            x = x[self.layer]
            x = self.model.ln_final(x)
@@ -612,6 +618,22 @@ class FrozenOpenCLIPEmbedder2(AbstractEmbModel):
                x = r(x, attn_mask=attn_mask)
        outputs["last"] = x.permute(1, 0, 2)  # LND -> NLD
        return outputs
    def text_transformer_forward_batch_first(self, x: torch.Tensor, attn_mask=None):
        x = x.permute(1, 0, 2) # LND -> NLD
        outputs = {}
        for i, r in enumerate(self.model.transformer.resblocks):
            if i == len(self.model.transformer.resblocks) - 1:
                outputs["penultimate"] = x
            if (
                self.model.transformer.grad_checkpointing
                and not torch.jit.is_scripting()
            ):
                x = checkpoint(r, x, attn_mask)
            else:
                x = r(x, attn_mask=attn_mask)
        outputs["last"] = x
        return outputs
    def encode(self, text):
        return self(text)
@@ -908,7 +930,7 @@ class SpatialRescaler(nn.Module):
            print(
                f"Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing."
            )
-            self.channel_mapper = nn.Conv2d(
+            self.channel_mapper = ops.Conv2d(
                in_channels,
                out_channels,
                kernel_size=kernel_size,
@@ -4,7 +4,7 @@ import os
 from functools import partial
 from inspect import isfunction
-import fsspec
+#import fsspec
 import numpy as np
 import torch
 from PIL import Image, ImageDraw, ImageFont
@@ -177,14 +177,17 @@ def instantiate_from_config(config):
 def get_obj_from_str(string, reload=False, invalidate_cache=True):
    package_directory_name = os.path.basename(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
    module, cls = string.rsplit(".", 1)
    if invalidate_cache:
        importlib.invalidate_caches()
    if reload:
        module_imp = importlib.import_module(module)
        importlib.reload(module_imp)
-    return getattr(importlib.import_module(module, package=package_directory_name), cls)
+    try:
        obj = getattr(importlib.import_module(module, package=os.path.basename(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))), cls)
    except:
        obj = getattr(importlib.import_module(module, package=os.path.dirname(os.path.dirname(os.path.abspath( __file__ )))), cls)
    return obj
 def append_zero(x):
Author	SHA1	Message	Date
Jukka Seppänen	99d49e912c	Update README.md	2026-04-29 18:30:56 +03:00
kijai	fe0d660f7c	Fix for latest ComfyUI	2026-03-15 17:02:02 +02:00
kijai	0613a9239f	version 1.0.3	2026-02-04 16:55:13 +02:00
kijai	4ebf22bc2c	Some cleanup	2026-02-04 16:53:20 +02:00
Jukka Seppänen	6e972b1f60	Merge pull request #174 from ComfyNodePRs/update-publish-yaml Update Github Action for Publishing to Comfy Registry	2026-02-04 16:49:40 +02:00
Jukka Seppänen	30aac9a766	Merge pull request #177 from hj-desperado/main Add dtype selection in decoder, aligning with encoder	2026-02-04 16:48:18 +02:00
kijai	159995e440	Fix for newer ComfyUI version	2026-02-04 16:47:23 +02:00
Jukka Seppänen	0311fb8e79	Merge pull request #199 from jorismak/feature/fix-sm12 fallback for xformers (fixes SM12)	2026-02-04 16:46:18 +02:00
Joris Mak	a9c2fc4115	cleanup try/catch location	2026-01-11 14:28:39 +01:00
Joris Mak	1caaaa50d9	Implement fallback for memory-efficient attention in xformer_attn_forward	2026-01-11 14:14:05 +01:00
kijai	c5cb1ce099	Update requirements.txt	2025-09-24 16:51:11 +03:00
kijai	6fe7a3f856	bump version	2025-06-07 12:17:40 +03:00
kijai	29f2e8be0e	Possible workaround for import issue in some configurations	2025-04-14 01:52:00 +03:00
jianghan01	a132c748ad	Add dtype selection in decoder, aligning with encoder	2025-04-02 17:04:19 +08:00
snomiao	d5b69963db	chore(publish): update GitHub Actions workflow for node publishing - Add permissions for issue writing - Update action version to v1 for publish-node-action - Add condition to run job only for 'kijai' repository owner	2025-01-20 21:36:40 +00:00
kijai	53fc4f82f1	clean some unnecessary dependencies	2024-07-07 02:44:51 +03:00
kijai	0597c9cdb6	Update pyproject.toml	2024-07-06 12:23:56 +03:00
kijai	59ca341c0a	workaround for open-clip-torch update fixes the attn_mask shape error	2024-07-06 12:22:30 +03:00
kijai	a8780cd188	use more comfy ops	2024-07-06 12:20:21 +03:00
Jukka Seppänen	2726b6ef34	Update requirements.txt	2024-07-06 11:52:45 +03:00
kijai	006754633a	Support SDXL latents directly Bypassing need for SUPIR VAE	2024-06-27 16:40:45 +03:00
Jukka Seppänen	c257cce555	Merge pull request #124 from haohaocreates/publish Add Github Action for Publishing to Comfy Registry	2024-06-27 12:29:20 +03:00
kijai	eda2cf504a	Update pyproject.toml	2024-06-27 12:28:18 +03:00
Jukka Seppänen	e1770b002a	Merge pull request #125 from haohaocreates/pyproject Add pyproject.toml for Custom Node Registry	2024-06-27 12:27:28 +03:00
haohaocreates	3320502310	Update pyproject.toml description	2024-05-21 10:56:33 -04:00
haohaocreates	91893f123d	chore(pyproject): Add pyproject.toml for Custom Node Registry	2024-05-21 14:46:39 +00:00
haohaocreates	1db91b1510	chore(publish): Add Github Action for Publishing to Comfy Registry	2024-05-21 14:46:38 +00:00
kijai	bbb2bd7e8f	Add version of the loader node to load CLIP separately	2024-05-20 11:12:07 +03:00
Kijai	5518e00a14	Remove redundant req	2024-04-23 13:04:10 +03:00
kijai	656e55e815	accelerate fixes	2024-04-14 11:27:17 +03:00
kijai	c97893850e	Make accelerate model loading target choseable	2024-04-14 02:01:54 +03:00
kijai	7596566a89	Accelerate	2024-04-13 17:48:47 +03:00
kijai	49afedaddd	Description for sampler	2024-04-08 21:10:46 +03:00
kijai	3a911c0cbd	Merge branch 'main' of https://github.com/kijai/ComfyUI-SUPIR	2024-04-07 02:43:25 +03:00
kijai	84e67594d0	Initial DESCRIPTIONs	2024-04-07 02:43:23 +03:00
Jukka Seppänen	472dbce4dc	Update README.md	2024-04-07 02:34:15 +03:00
kijai	d20c5190a1	Update nodes_v2.py	2024-04-05 10:59:10 +03:00
kijai	30f60a3b03	clip decode on gpu	2024-04-05 10:51:32 +03:00
Jukka Seppänen	e62f3c5dac	Create LICENSE	2024-03-30 00:29:23 +02:00
Jukka Seppänen	91ecd6f8de	Delete LICENSE	2024-03-30 00:28:57 +02:00
Jukka Seppänen	ba7fceeaac	Update LICENSE	2024-03-30 00:28:33 +02:00
kijai	2ce3ced581	Tiled sampler consequential runs workaround Copy the conds so they are not permanently modified by the Tiled samplers	2024-03-26 10:34:17 +02:00
Kijai	3f340a6b96	update example	2024-03-25 18:12:55 +02:00
Kijai	203bc93933	Update nodes_v2.py	2024-03-25 18:06:11 +02:00
Kijai	eb7d1e31ae	Update nodes_v2.py	2024-03-25 17:54:24 +02:00
Kijai	61cbb153e6	Update nodes_v2.py	2024-03-25 17:47:48 +02:00
kijai	5c54bb1437	Better xformers check Now relies in the comfy xformers check and can thus be disabled with the command line argument --disable-xformers	2024-03-25 10:46:36 +02:00
kijai	2cd64cdfce	cleanup	2024-03-25 10:17:45 +02:00
kijai	80d20731a4	Update nodes_v2.py	2024-03-25 01:29:31 +02:00
kijai	dced82b070	Smarter model loading and support comfy latents - Less spiky memory usage when using v2 model loader node - Support latents from other comfy nodes, when using latents from SUPIR nodes the behaviour is same as before: fully random noise	2024-03-24 15:45:10 +02:00
kijai	e933093166	fix decoder dtype	2024-03-22 08:48:48 +02:00
kijai	fb6f99e7d5	clip decoder output	2024-03-22 08:37:04 +02:00
Jukka Seppänen	4a10dd4da5	Update README.md	2024-03-22 00:40:05 +02:00
kijai	adc82313eb	Make model loader node without sdxl_model selection for clarity	2024-03-22 00:37:21 +02:00
Kijai	61942e250c	Support using comfy model loading, including loras These inputs are optional, when used the sdxl_model dropdown is ignored, I don't want to remove it as of yet as it would break old workflows.	2024-03-21 16:06:08 +02:00
Kijai	98291ac559	Update nodes_v2.py	2024-03-20 18:45:46 +02:00
Kijai	d7ab37bed0	Update nodes_v2.py	2024-03-20 18:42:47 +02:00
Kijai	14480f926a	Default to fp16 unet with 'auto'	2024-03-19 16:25:54 +02:00
kijai	dbc22b88fb	batch fixes	2024-03-19 11:02:54 +02:00
kijai	734054ea64	Update nodes_v2.py	2024-03-19 10:19:32 +02:00
Kijai	e460953d7c	Fix resize and batch captioning	2024-03-18 17:44:14 +02:00
Kijai	033f83ad17	Batch support for video	2024-03-18 16:49:57 +02:00
kijai	2bd6421b6f	Flip cfg linear scale behaviour for v2 sampler Makes more sense to me, scales cfg linearly from cfg_scale_start at first step to cfg_scale_end at last	2024-03-18 01:48:19 +02:00
kijai	b887b9e4be	clean prints	2024-03-17 17:55:24 +02:00
kijai	166b5920d3	fix rescaling (again)	2024-03-17 17:50:58 +02:00
kijai	714256e04a	Merge branch 'main' of https://github.com/kijai/ComfyUI-SUPIR	2024-03-17 16:39:22 +02:00
kijai	092c6b81dc	ignore eta setting for EDM samplers	2024-03-17 16:39:20 +02:00
Jukka Seppänen	a766cb0a9f	Update README.md	2024-03-17 15:36:21 +02:00
kijai	747f07b99d	Big update: Add v2 nodes Separate the process to multiple nodes for efficiency and better understanding of the process.	2024-03-17 14:03:11 +02:00
kijai	2d1ecb1bf0	Revert resize method	2024-03-11 23:12:00 +02:00
kijai	bda46b35b5	fix autoresize	2024-03-11 21:56:43 +02:00
kijai	8341b91e66	Update readme	2024-03-11 09:44:00 +02:00
kijai	8a4beaa547	Fix tiled sampling	2024-03-11 09:31:27 +02:00
kijai	aadedb1d29	clean up	2024-03-10 14:59:56 +02:00
kijai	c1e8bc5f80	Add RestoreDPMPP2MSampler Better suited for lightning models	2024-03-10 14:25:50 +02:00
kijai	e0324bed08	Merge branch 'main' of https://github.com/kijai/ComfyUI-SUPIR	2024-03-09 17:15:20 +02:00
kijai	24103daee0	Update nodes.py	2024-03-09 17:15:12 +02:00
Jukka Seppänen	c5d1e1bebc	Update README.md	2024-03-09 17:08:25 +02:00
Jukka Seppänen	78e947e7db	Update README.md	2024-03-09 17:07:25 +02:00
kijai	35ba391507	Choice to use fp8 for unet and vae Experimental, seems to work for saving memory, but there's a quality hit	2024-03-09 16:58:59 +02:00
kijai	763ab95734	Add some cpu offloading	2024-03-09 12:53:50 +02:00