docs: sync with converter consolidation

Docs described the pre-#67 suite: a removed standalone LCM converter node, a removed model_version input, a nonexistent CoreMLDetailerHookProvider node, and a deleted lcm/converter.py file. - drop LCM converter node docs; LCM checkpoints are auto-detected by the consolidated CoreMLConverter - remove model_version input and invented 512-768 resolution range - replace phantom detailer-hook fix with real workarounds - fix Python support claim (3.12+ per requires-python) - update LCM support-matrix row, drop stale line reference Refs #67, #68
docs: rewrite README and split into docs/ pages
2026-07-09 18:33:17 +02:00 · 2026-07-09 18:30:26 +02:00 · 2026-06-13 15:48:42 +02:00 · 2026-06-13 14:11:20 +02:00 · 2026-05-27 05:56:45 +02:00 · 2026-05-27 02:54:12 +02:00
37 changed files with 1644 additions and 3058 deletions
@@ -1,23 +0,0 @@
 name: Tier 1 — Smoke (macOS self-hosted)
 # macOS smoke tests run on the self-hosted Apple Silicon runner instead of
 # GitHub-hosted macOS (10x minute multiplier), which exhausts the included
 # Actions minutes too quickly.
 on:
  push:
    branches: [main]
  pull_request:
 jobs:
  smoke:
    runs-on: [self-hosted, macOS, ARM64, coreml]
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v4
      # The self-hosted runner provides uv; no setup-uv action needed.
      - name: uv sync
        run: uv sync --no-install-project
      - name: Run Tier 1 (synthetic micro-UNet smoke)
        run: uv run pytest -m smoke tests/ -v
@@ -3,3 +3,6 @@ __pycache__/
 models/
 .venv/
 test_results/
 *.log
 .DS_Store
 .claude/
@@ -0,0 +1 @@
 3.12
@@ -1,674 +1,21 @@
-                    GNU GENERAL PUBLIC LICENSE
+MIT License
-                       Version 3, 29 June 2007
+
-
+Copyright (c) 2023-2026 Adrian Szczepański
- Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+
- Everyone is permitted to copy and distribute verbatim copies
+Permission is hereby granted, free of charge, to any person obtaining a copy
- of this license document, but changing it is not allowed.
+of this software and associated documentation files (the "Software"), to deal
-
+in the Software without restriction, including without limitation the rights
-                            Preamble
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-
+copies of the Software, and to permit persons to whom the Software is
-  The GNU General Public License is a free, copyleft license for
+furnished to do so, subject to the following conditions:
-software and other kinds of works.
+
-
+The above copyright notice and this permission notice shall be included in all
-  The licenses for most software and other practical works are designed
+copies or substantial portions of the Software.
-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
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-share and change all versions of a program--to make sure it remains free
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-software for all its users.  We, the Free Software Foundation, use the
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-GNU General Public License for most of our software; it applies also to
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-any other work released this way by its authors.  You can apply it to
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-your programs, too.
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-
+SOFTWARE.
  When we speak of free software, we are referring to freedom, not
 price.  Our General Public Licenses are designed to make sure that you
 have the freedom to distribute copies of free software (and charge for
 them if you wish), that you receive source code or can get it if you
 want it, that you can change the software or use pieces of it in new
 free programs, and that you know you can do these things.
  To protect your rights, we need to prevent others from denying you
 these rights or asking you to surrender the rights.  Therefore, you have
 certain responsibilities if you distribute copies of the software, or if
 you modify it: responsibilities to respect the freedom of others.
  For example, if you distribute copies of such a program, whether
 gratis or for a fee, you must pass on to the recipients the same
 freedoms that you received.  You must make sure that they, too, receive
 or can get the source code.  And you must show them these terms so they
 know their rights.
  Developers that use the GNU GPL protect your rights with two steps:
 (1) assert copyright on the software, and (2) offer you this License
 giving you legal permission to copy, distribute and/or modify it.
  For the developers' and authors' protection, the GPL clearly explains
 that there is no warranty for this free software.  For both users' and
 authors' sake, the GPL requires that modified versions be marked as
 changed, so that their problems will not be attributed erroneously to
 authors of previous versions.
  Some devices are designed to deny users access to install or run
 modified versions of the software inside them, although the manufacturer
 can do so.  This is fundamentally incompatible with the aim of
 protecting users' freedom to change the software.  The systematic
 pattern of such abuse occurs in the area of products for individuals to
 use, which is precisely where it is most unacceptable.  Therefore, we
 have designed this version of the GPL to prohibit the practice for those
 products.  If such problems arise substantially in other domains, we
 stand ready to extend this provision to those domains in future versions
 of the GPL, as needed to protect the freedom of users.
  Finally, every program is threatened constantly by software patents.
 States should not allow patents to restrict development and use of
 software on general-purpose computers, but in those that do, we wish to
 avoid the special danger that patents applied to a free program could
 make it effectively proprietary.  To prevent this, the GPL assures that
 patents cannot be used to render the program non-free.
  The precise terms and conditions for copying, distribution and
 modification follow.
                       TERMS AND CONDITIONS
  0. Definitions.
  "This License" refers to version 3 of the GNU General Public License.
  "Copyright" also means copyright-like laws that apply to other kinds of
 works, such as semiconductor masks.
  "The Program" refers to any copyrightable work licensed under this
 License.  Each licensee is addressed as "you".  "Licensees" and
 "recipients" may be individuals or organizations.
  To "modify" a work means to copy from or adapt all or part of the work
 in a fashion requiring copyright permission, other than the making of an
 exact copy.  The resulting work is called a "modified version" of the
 earlier work or a work "based on" the earlier work.
  A "covered work" means either the unmodified Program or a work based
 on the Program.
  To "propagate" a work means to do anything with it that, without
 permission, would make you directly or secondarily liable for
 infringement under applicable copyright law, except executing it on a
 computer or modifying a private copy.  Propagation includes copying,
 distribution (with or without modification), making available to the
 public, and in some countries other activities as well.
  To "convey" a work means any kind of propagation that enables other
 parties to make or receive copies.  Mere interaction with a user through
 a computer network, with no transfer of a copy, is not conveying.
  An interactive user interface displays "Appropriate Legal Notices"
 to the extent that it includes a convenient and prominently visible
 feature that (1) displays an appropriate copyright notice, and (2)
 tells the user that there is no warranty for the work (except to the
 extent that warranties are provided), that licensees may convey the
 work under this License, and how to view a copy of this License.  If
 the interface presents a list of user commands or options, such as a
 menu, a prominent item in the list meets this criterion.
  1. Source Code.
  The "source code" for a work means the preferred form of the work
 for making modifications to it.  "Object code" means any non-source
 form of a work.
  A "Standard Interface" means an interface that either is an official
 standard defined by a recognized standards body, or, in the case of
 interfaces specified for a particular programming language, one that
 is widely used among developers working in that language.
  The "System Libraries" of an executable work include anything, other
 than the work as a whole, that (a) is included in the normal form of
 packaging a Major Component, but which is not part of that Major
 Component, and (b) serves only to enable use of the work with that
 Major Component, or to implement a Standard Interface for which an
 implementation is available to the public in source code form.  A
 "Major Component", in this context, means a major essential component
 (kernel, window system, and so on) of the specific operating system
 (if any) on which the executable work runs, or a compiler used to
 produce the work, or an object code interpreter used to run it.
  The "Corresponding Source" for a work in object code form means all
 the source code needed to generate, install, and (for an executable
 work) run the object code and to modify the work, including scripts to
 control those activities.  However, it does not include the work's
 System Libraries, or general-purpose tools or generally available free
 programs which are used unmodified in performing those activities but
 which are not part of the work.  For example, Corresponding Source
 includes interface definition files associated with source files for
 the work, and the source code for shared libraries and dynamically
 linked subprograms that the work is specifically designed to require,
 such as by intimate data communication or control flow between those
 subprograms and other parts of the work.
  The Corresponding Source need not include anything that users
 can regenerate automatically from other parts of the Corresponding
 Source.
  The Corresponding Source for a work in source code form is that
 same work.
  2. Basic Permissions.
  All rights granted under this License are granted for the term of
 copyright on the Program, and are irrevocable provided the stated
 conditions are met.  This License explicitly affirms your unlimited
 permission to run the unmodified Program.  The output from running a
 covered work is covered by this License only if the output, given its
 content, constitutes a covered work.  This License acknowledges your
 rights of fair use or other equivalent, as provided by copyright law.
  You may make, run and propagate covered works that you do not
 convey, without conditions so long as your license otherwise remains
 in force.  You may convey covered works to others for the sole purpose
 of having them make modifications exclusively for you, or provide you
 with facilities for running those works, provided that you comply with
 the terms of this License in conveying all material for which you do
 not control copyright.  Those thus making or running the covered works
 for you must do so exclusively on your behalf, under your direction
 and control, on terms that prohibit them from making any copies of
 your copyrighted material outside their relationship with you.
  Conveying under any other circumstances is permitted solely under
 the conditions stated below.  Sublicensing is not allowed; section 10
 makes it unnecessary.
  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
  No covered work shall be deemed part of an effective technological
 measure under any applicable law fulfilling obligations under article
 11 of the WIPO copyright treaty adopted on 20 December 1996, or
 similar laws prohibiting or restricting circumvention of such
 measures.
  When you convey a covered work, you waive any legal power to forbid
 circumvention of technological measures to the extent such circumvention
 is effected by exercising rights under this License with respect to
 the covered work, and you disclaim any intention to limit operation or
 modification of the work as a means of enforcing, against the work's
 users, your or third parties' legal rights to forbid circumvention of
 technological measures.
  4. Conveying Verbatim Copies.
  You may convey verbatim copies of the Program's source code as you
 receive it, in any medium, provided that you conspicuously and
 appropriately publish on each copy an appropriate copyright notice;
 keep intact all notices stating that this License and any
 non-permissive terms added in accord with section 7 apply to the code;
 keep intact all notices of the absence of any warranty; and give all
 recipients a copy of this License along with the Program.
  You may charge any price or no price for each copy that you convey,
 and you may offer support or warranty protection for a fee.
  5. Conveying Modified Source Versions.
  You may convey a work based on the Program, or the modifications to
 produce it from the Program, in the form of source code under the
 terms of section 4, provided that you also meet all of these conditions:
    a) The work must carry prominent notices stating that you modified
    it, and giving a relevant date.
    b) The work must carry prominent notices stating that it is
    released under this License and any conditions added under section
    7.  This requirement modifies the requirement in section 4 to
    "keep intact all notices".
    c) You must license the entire work, as a whole, under this
    License to anyone who comes into possession of a copy.  This
    License will therefore apply, along with any applicable section 7
    additional terms, to the whole of the work, and all its parts,
    regardless of how they are packaged.  This License gives no
    permission to license the work in any other way, but it does not
    invalidate such permission if you have separately received it.
    d) If the work has interactive user interfaces, each must display
    Appropriate Legal Notices; however, if the Program has interactive
    interfaces that do not display Appropriate Legal Notices, your
    work need not make them do so.
  A compilation of a covered work with other separate and independent
 works, which are not by their nature extensions of the covered work,
 and which are not combined with it such as to form a larger program,
 in or on a volume of a storage or distribution medium, is called an
 "aggregate" if the compilation and its resulting copyright are not
 used to limit the access or legal rights of the compilation's users
 beyond what the individual works permit.  Inclusion of a covered work
 in an aggregate does not cause this License to apply to the other
 parts of the aggregate.
  6. Conveying Non-Source Forms.
  You may convey a covered work in object code form under the terms
 of sections 4 and 5, provided that you also convey the
 machine-readable Corresponding Source under the terms of this License,
 in one of these ways:
    a) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by the
    Corresponding Source fixed on a durable physical medium
    customarily used for software interchange.
    b) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by a
    written offer, valid for at least three years and valid for as
    long as you offer spare parts or customer support for that product
    model, to give anyone who possesses the object code either (1) a
    copy of the Corresponding Source for all the software in the
    product that is covered by this License, on a durable physical
    medium customarily used for software interchange, for a price no
    more than your reasonable cost of physically performing this
    conveying of source, or (2) access to copy the
    Corresponding Source from a network server at no charge.
    c) Convey individual copies of the object code with a copy of the
    written offer to provide the Corresponding Source.  This
    alternative is allowed only occasionally and noncommercially, and
    only if you received the object code with such an offer, in accord
    with subsection 6b.
    d) Convey the object code by offering access from a designated
    place (gratis or for a charge), and offer equivalent access to the
    Corresponding Source in the same way through the same place at no
    further charge.  You need not require recipients to copy the
    Corresponding Source along with the object code.  If the place to
    copy the object code is a network server, the Corresponding Source
    may be on a different server (operated by you or a third party)
    that supports equivalent copying facilities, provided you maintain
    clear directions next to the object code saying where to find the
    Corresponding Source.  Regardless of what server hosts the
    Corresponding Source, you remain obligated to ensure that it is
    available for as long as needed to satisfy these requirements.
    e) Convey the object code using peer-to-peer transmission, provided
    you inform other peers where the object code and Corresponding
    Source of the work are being offered to the general public at no
    charge under subsection 6d.
  A separable portion of the object code, whose source code is excluded
 from the Corresponding Source as a System Library, need not be
 included in conveying the object code work.
  A "User Product" is either (1) a "consumer product", which means any
 tangible personal property which is normally used for personal, family,
 or household purposes, or (2) anything designed or sold for incorporation
 into a dwelling.  In determining whether a product is a consumer product,
 doubtful cases shall be resolved in favor of coverage.  For a particular
 product received by a particular user, "normally used" refers to a
 typical or common use of that class of product, regardless of the status
 of the particular user or of the way in which the particular user
 actually uses, or expects or is expected to use, the product.  A product
 is a consumer product regardless of whether the product has substantial
 commercial, industrial or non-consumer uses, unless such uses represent
 the only significant mode of use of the product.
  "Installation Information" for a User Product means any methods,
 procedures, authorization keys, or other information required to install
 and execute modified versions of a covered work in that User Product from
 a modified version of its Corresponding Source.  The information must
 suffice to ensure that the continued functioning of the modified object
 code is in no case prevented or interfered with solely because
 modification has been made.
  If you convey an object code work under this section in, or with, or
 specifically for use in, a User Product, and the conveying occurs as
 part of a transaction in which the right of possession and use of the
 User Product is transferred to the recipient in perpetuity or for a
 fixed term (regardless of how the transaction is characterized), the
 Corresponding Source conveyed under this section must be accompanied
 by the Installation Information.  But this requirement does not apply
 if neither you nor any third party retains the ability to install
 modified object code on the User Product (for example, the work has
 been installed in ROM).
  The requirement to provide Installation Information does not include a
 requirement to continue to provide support service, warranty, or updates
 for a work that has been modified or installed by the recipient, or for
 the User Product in which it has been modified or installed.  Access to a
 network may be denied when the modification itself materially and
 adversely affects the operation of the network or violates the rules and
 protocols for communication across the network.
  Corresponding Source conveyed, and Installation Information provided,
 in accord with this section must be in a format that is publicly
 documented (and with an implementation available to the public in
 source code form), and must require no special password or key for
 unpacking, reading or copying.
  7. Additional Terms.
  "Additional permissions" are terms that supplement the terms of this
 License by making exceptions from one or more of its conditions.
 Additional permissions that are applicable to the entire Program shall
 be treated as though they were included in this License, to the extent
 that they are valid under applicable law.  If additional permissions
 apply only to part of the Program, that part may be used separately
 under those permissions, but the entire Program remains governed by
 this License without regard to the additional permissions.
  When you convey a copy of a covered work, you may at your option
 remove any additional permissions from that copy, or from any part of
 it.  (Additional permissions may be written to require their own
 removal in certain cases when you modify the work.)  You may place
 additional permissions on material, added by you to a covered work,
 for which you have or can give appropriate copyright permission.
  Notwithstanding any other provision of this License, for material you
 add to a covered work, you may (if authorized by the copyright holders of
 that material) supplement the terms of this License with terms:
    a) Disclaiming warranty or limiting liability differently from the
    terms of sections 15 and 16 of this License; or
    b) Requiring preservation of specified reasonable legal notices or
    author attributions in that material or in the Appropriate Legal
    Notices displayed by works containing it; or
    c) Prohibiting misrepresentation of the origin of that material, or
    requiring that modified versions of such material be marked in
    reasonable ways as different from the original version; or
    d) Limiting the use for publicity purposes of names of licensors or
    authors of the material; or
    e) Declining to grant rights under trademark law for use of some
    trade names, trademarks, or service marks; or
    f) Requiring indemnification of licensors and authors of that
    material by anyone who conveys the material (or modified versions of
    it) with contractual assumptions of liability to the recipient, for
    any liability that these contractual assumptions directly impose on
    those licensors and authors.
  All other non-permissive additional terms are considered "further
 restrictions" within the meaning of section 10.  If the Program as you
 received it, or any part of it, contains a notice stating that it is
 governed by this License along with a term that is a further
 restriction, you may remove that term.  If a license document contains
 a further restriction but permits relicensing or conveying under this
 License, you may add to a covered work material governed by the terms
 of that license document, provided that the further restriction does
 not survive such relicensing or conveying.
  If you add terms to a covered work in accord with this section, you
 must place, in the relevant source files, a statement of the
 additional terms that apply to those files, or a notice indicating
 where to find the applicable terms.
  Additional terms, permissive or non-permissive, may be stated in the
 form of a separately written license, or stated as exceptions;
 the above requirements apply either way.
  8. Termination.
  You may not propagate or modify a covered work except as expressly
 provided under this License.  Any attempt otherwise to propagate or
 modify it is void, and will automatically terminate your rights under
 this License (including any patent licenses granted under the third
 paragraph of section 11).
  However, if you cease all violation of this License, then your
 license from a particular copyright holder is reinstated (a)
 provisionally, unless and until the copyright holder explicitly and
 finally terminates your license, and (b) permanently, if the copyright
 holder fails to notify you of the violation by some reasonable means
 prior to 60 days after the cessation.
  Moreover, your license from a particular copyright holder is
 reinstated permanently if the copyright holder notifies you of the
 violation by some reasonable means, this is the first time you have
 received notice of violation of this License (for any work) from that
 copyright holder, and you cure the violation prior to 30 days after
 your receipt of the notice.
  Termination of your rights under this section does not terminate the
 licenses of parties who have received copies or rights from you under
 this License.  If your rights have been terminated and not permanently
 reinstated, you do not qualify to receive new licenses for the same
 material under section 10.
  9. Acceptance Not Required for Having Copies.
  You are not required to accept this License in order to receive or
 run a copy of the Program.  Ancillary propagation of a covered work
 occurring solely as a consequence of using peer-to-peer transmission
 to receive a copy likewise does not require acceptance.  However,
 nothing other than this License grants you permission to propagate or
 modify any covered work.  These actions infringe copyright if you do
 not accept this License.  Therefore, by modifying or propagating a
 covered work, you indicate your acceptance of this License to do so.
  10. Automatic Licensing of Downstream Recipients.
  Each time you convey a covered work, the recipient automatically
 receives a license from the original licensors, to run, modify and
 propagate that work, subject to this License.  You are not responsible
 for enforcing compliance by third parties with this License.
  An "entity transaction" is a transaction transferring control of an
 organization, or substantially all assets of one, or subdividing an
 organization, or merging organizations.  If propagation of a covered
 work results from an entity transaction, each party to that
 transaction who receives a copy of the work also receives whatever
 licenses to the work the party's predecessor in interest had or could
 give under the previous paragraph, plus a right to possession of the
 Corresponding Source of the work from the predecessor in interest, if
 the predecessor has it or can get it with reasonable efforts.
  You may not impose any further restrictions on the exercise of the
 rights granted or affirmed under this License.  For example, you may
 not impose a license fee, royalty, or other charge for exercise of
 rights granted under this License, and you may not initiate litigation
 (including a cross-claim or counterclaim in a lawsuit) alleging that
 any patent claim is infringed by making, using, selling, offering for
 sale, or importing the Program or any portion of it.
  11. Patents.
  A "contributor" is a copyright holder who authorizes use under this
 License of the Program or a work on which the Program is based.  The
 work thus licensed is called the contributor's "contributor version".
  A contributor's "essential patent claims" are all patent claims
 owned or controlled by the contributor, whether already acquired or
 hereafter acquired, that would be infringed by some manner, permitted
 by this License, of making, using, or selling its contributor version,
 but do not include claims that would be infringed only as a
 consequence of further modification of the contributor version.  For
 purposes of this definition, "control" includes the right to grant
 patent sublicenses in a manner consistent with the requirements of
 this License.
  Each contributor grants you a non-exclusive, worldwide, royalty-free
 patent license under the contributor's essential patent claims, to
 make, use, sell, offer for sale, import and otherwise run, modify and
 propagate the contents of its contributor version.
  In the following three paragraphs, a "patent license" is any express
 agreement or commitment, however denominated, not to enforce a patent
 (such as an express permission to practice a patent or covenant not to
 sue for patent infringement).  To "grant" such a patent license to a
 party means to make such an agreement or commitment not to enforce a
 patent against the party.
  If you convey a covered work, knowingly relying on a patent license,
 and the Corresponding Source of the work is not available for anyone
 to copy, free of charge and under the terms of this License, through a
 publicly available network server or other readily accessible means,
 then you must either (1) cause the Corresponding Source to be so
 available, or (2) arrange to deprive yourself of the benefit of the
 patent license for this particular work, or (3) arrange, in a manner
 consistent with the requirements of this License, to extend the patent
 license to downstream recipients.  "Knowingly relying" means you have
 actual knowledge that, but for the patent license, your conveying the
 covered work in a country, or your recipient's use of the covered work
 in a country, would infringe one or more identifiable patents in that
 country that you have reason to believe are valid.
  If, pursuant to or in connection with a single transaction or
 arrangement, you convey, or propagate by procuring conveyance of, a
 covered work, and grant a patent license to some of the parties
 receiving the covered work authorizing them to use, propagate, modify
 or convey a specific copy of the covered work, then the patent license
 you grant is automatically extended to all recipients of the covered
 work and works based on it.
  A patent license is "discriminatory" if it does not include within
 the scope of its coverage, prohibits the exercise of, or is
 conditioned on the non-exercise of one or more of the rights that are
 specifically granted under this License.  You may not convey a covered
 work if you are a party to an arrangement with a third party that is
 in the business of distributing software, under which you make payment
 to the third party based on the extent of your activity of conveying
 the work, and under which the third party grants, to any of the
 parties who would receive the covered work from you, a discriminatory
 patent license (a) in connection with copies of the covered work
 conveyed by you (or copies made from those copies), or (b) primarily
 for and in connection with specific products or compilations that
 contain the covered work, unless you entered into that arrangement,
 or that patent license was granted, prior to 28 March 2007.
  Nothing in this License shall be construed as excluding or limiting
 any implied license or other defenses to infringement that may
 otherwise be available to you under applicable patent law.
  12. No Surrender of Others' Freedom.
  If conditions are imposed on you (whether by court order, agreement or
 otherwise) that contradict the conditions of this License, they do not
 excuse you from the conditions of this License.  If you cannot convey a
 covered work so as to satisfy simultaneously your obligations under this
 License and any other pertinent obligations, then as a consequence you may
 not convey it at all.  For example, if you agree to terms that obligate you
 to collect a royalty for further conveying from those to whom you convey
 the Program, the only way you could satisfy both those terms and this
 License would be to refrain entirely from conveying the Program.
  13. Use with the GNU Affero General Public License.
  Notwithstanding any other provision of this License, you have
 permission to link or combine any covered work with a work licensed
 under version 3 of the GNU Affero General Public License into a single
 combined work, and to convey the resulting work.  The terms of this
 License will continue to apply to the part which is the covered work,
 but the special requirements of the GNU Affero General Public License,
 section 13, concerning interaction through a network will apply to the
 combination as such.
  14. Revised Versions of this License.
  The Free Software Foundation may publish revised and/or new versions of
 the GNU General Public License from time to time.  Such new versions will
 be similar in spirit to the present version, but may differ in detail to
 address new problems or concerns.
  Each version is given a distinguishing version number.  If the
 Program specifies that a certain numbered version of the GNU General
 Public License "or any later version" applies to it, you have the
 option of following the terms and conditions either of that numbered
 version or of any later version published by the Free Software
 Foundation.  If the Program does not specify a version number of the
 GNU General Public License, you may choose any version ever published
 by the Free Software Foundation.
  If the Program specifies that a proxy can decide which future
 versions of the GNU General Public License can be used, that proxy's
 public statement of acceptance of a version permanently authorizes you
 to choose that version for the Program.
  Later license versions may give you additional or different
 permissions.  However, no additional obligations are imposed on any
 author or copyright holder as a result of your choosing to follow a
 later version.
  15. Disclaimer of Warranty.
  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
 APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
 HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
 OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
 THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
 IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
 ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
  16. Limitation of Liability.
  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
 WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
 THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
 GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
 USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
 DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
 PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
 EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
 SUCH DAMAGES.
  17. Interpretation of Sections 15 and 16.
  If the disclaimer of warranty and limitation of liability provided
 above cannot be given local legal effect according to their terms,
 reviewing courts shall apply local law that most closely approximates
 an absolute waiver of all civil liability in connection with the
 Program, unless a warranty or assumption of liability accompanies a
 copy of the Program in return for a fee.
                     END OF TERMS AND CONDITIONS
            How to Apply These Terms to Your New Programs
  If you develop a new program, and you want it to be of the greatest
 possible use to the public, the best way to achieve this is to make it
 free software which everyone can redistribute and change under these terms.
  To do so, attach the following notices to the program.  It is safest
 to attach them to the start of each source file to most effectively
 state the exclusion of warranty; and each file should have at least
 the "copyright" line and a pointer to where the full notice is found.
    <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>.
@@ -1,452 +1,113 @@
 # Core ML Suite for ComfyUI
-## Overview
+Custom nodes for [ComfyUI](https://github.com/comfyanonymous/ComfyUI) that run
 Stable Diffusion UNets as [Core ML](https://developer.apple.com/documentation/coreml)
 models on Apple Silicon (M1/M2/M3). Core ML can use the Apple Neural Engine
 (ANE), which is unavailable to PyTorch — on an M2 Pro 32 GB, SD1.5 at 512×512
 generates roughly **1.5–2× faster** than the standard PyTorch/MPS path.
-Welcome! In this repository you'll find a set of custom nodes for [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
+You convert a Stable Diffusion checkpoint to a Core ML model with the nodes in
-that allows you to use Core ML models in your ComfyUI workflows.
+this suite, then sample from it like any other ComfyUI workflow.
 These models are designed to leverage the Apple Neural Engine (ANE) on Apple Silicon (M1/M2) machines,
 thereby enhancing your workflows and improving performance.
-If you're not sure how to obtain these models, you can download them
+> [!IMPORTANT]
-[here](https://huggingface.co/coreml-community) or convert your own checkpoints
+> **Convert your own checkpoints — that is the only supported path.** This
-directly with the conversion nodes in this suite (see [How to use](#how-to-use)).
+> suite uses its own input dimensions, naming convention, and metadata
 > (produced by the [coreml-diffusion](https://github.com/aszc-dev/coreml-diffusion)
 > package). Pre-converted Core ML models from elsewhere (e.g. the
 > coreml-community Hugging Face org) are **not** supported. Conversion is cheap
 > and runs on your machine, so there is no need to download Core ML models.
-In simple terms, think of Core ML models as a tool that can help your ComfyUI work faster and more efficiently.
+## Installation
 For instance, during my tests on an M2 Pro 32GB machine,
 the use of Core ML models sped up the generation of 512x512 images by a factor
 of approximately 1.5 to 2 times.
-## Getting Started
+### ComfyUI-Manager (recommended)
-To start using custom nodes in your ComfyUI, follow these simple steps:
+Open **Manager → Install Custom Nodes**, search for `Core ML`, click
 **Install**, and restart ComfyUI.
-1. Clone or download this repository: You can do this directly into the custom_nodes directory of your ComfyUI.
+### Manual
 2. Install the dependencies: You'll need to use a package manager like pip to do this.
-That's it! You're now ready to start enhancing your ComfyUI workflows with Core ML models.
+```bash
 cd /path/to/comfyui/custom_nodes
 git clone https://github.com/aszc-dev/ComfyUI-CoreMLSuite.git
 cd ComfyUI-CoreMLSuite
 pip install -r requirements.txt
 ```
- Check [Installation](#installation) for more details on installation.
+Dependencies (`coreml-diffusion`, `coremltools`, `numpy`, `diffusers`) install
- Check [How to use](#how-to-use) for more details on how to use the custom nodes.
+from PyPI. PyTorch is intentionally **not** pinned — it is provided by your
- Check [Example Workflows](#example-workflows) for some example workflows.
+ComfyUI host, and a hard cap here would downgrade it and break ComfyUI.
 ## Quickstart
 1. Put a SD1.5 checkpoint in `models/checkpoints`.
 2. Add the **Convert Checkpoint to Core ML** node, select the checkpoint, and
   queue once. It writes a `.mlpackage` to `models/unet` (cached by name — it
   won't reconvert next time).
 3. Sample with the **Core ML Sampler** node, decoding the latent with a normal
   VAE Decode. CLIP and VAE come from standard ComfyUI nodes.
 See [docs/workflows.md](docs/workflows.md) for complete example graphs (txt2img,
 ControlNet, LoRA, LCM, SDXL).
 ## Which compute unit should I pick?
 The **compute unit** selects the hardware Core ML runs on. Pair it with the
 attention implementation chosen at conversion time:
 | Model | Convert with | Load with | Runs on |
 |---|---|---|---|
 | SD1.5 @ 512×512 | `SPLIT_EINSUM` | `CPU_AND_NE` | Neural Engine (fastest) |
 | SD1.5 @ larger sizes | `ORIGINAL` | `CPU_AND_GPU` | GPU |
 | SDXL | `ORIGINAL` | `CPU_AND_GPU` | GPU (ANE unsupported) |
 `CPU_AND_NE` is usually the fastest option for SD1.5 — often faster than `ALL`.
 This suite uses Core ML compute units only; it never touches PyTorch MPS, so
 `PYTORCH_ENABLE_MPS_FALLBACK` is irrelevant to these nodes. Full reasoning and
 benchmarks: [docs/hardware.md](docs/hardware.md).
 ## Documentation
 - [Hardware & compute units](docs/hardware.md) — ANE vs GPU vs MPS, attention
  implementations, which to choose.
 - [Nodes](docs/nodes.md) — full reference for every node.
 - [Conversion](docs/conversion.md) — how conversion works, caching,
  quantization.
 - [Example workflows](docs/workflows.md) — annotated example graphs.
 - [FAQ](docs/faq.md) — answers to common questions.
 - [Troubleshooting](docs/troubleshooting.md) — common errors and fixes.
 - [Limitations & support matrix](docs/limitations.md) — what is and isn't
  supported.
 ## Glossary
- **Core ML**: A machine learning framework developed by Apple. It's used to run machine learning models on Apple
+- **Core ML** — Apple's on-device machine-learning framework.
-  devices.
+- **`.mlpackage`** — the Core ML model format this suite produces and loads.
- **Core ML Model**: A machine learning model that can be run on Apple devices using Core ML.
+- **ANE** — Apple Neural Engine, a hardware accelerator for ML.
- **mlmodelc**: A compiled Core ML model. This is the recommended format for Core ML models.
+- **Compute unit** — which hardware Core ML uses (`CPU_AND_NE`, `CPU_AND_GPU`,
- **mlpackage**: A Core ML model packaged in a directory. This is the default format for Core ML models.
+  `CPU_ONLY`, `ALL`).
- **ANE**: Apple Neural Engine. A hardware accelerator for machine learning tasks on Apple devices.
+- **Attention implementation** — `SPLIT_EINSUM` / `SPLIT_EINSUM_V2` (ANE-friendly)
- **Compute Unit**: A Core ML option that allows you to specify the hardware on which the model should run.
+  or `ORIGINAL` (GPU-friendly), chosen at conversion.
    - **CPU_AND_ANE**: A Core ML compute unit option that allows the model to run on both the CPU and ANE. This is the
      default option.
    - **CPU_AND_GPU**: A Core ML compute unit option that allows the model to run on both the CPU and GPU.
    - **CPU_ONLY**: A Core ML compute unit option that allows the model to run on the CPU only.
    - **ALL**: A Core ML compute unit option that allows the model to run on all available hardware.
 - **CLIP**: Contrastive Language-Image Pre-training. A model that learns visual concepts from natural language
  supervision. It's used as a text encoder in Stable Diffusion.
 - **VAE**: Variational Autoencoder. A model that learns a latent representation of images. It's used as a prior in
  Stable Diffusion.
 - **Checkpoint**: A file that contains the weights of a model. It's used to load models in Stable Diffusion.
 - **LCM**: [Latent Consistency Model](https://latent-consistency-models.github.io/). A type of model designed to
  generate images with as few steps as possible.
 > [!NOTE]
 > Note on Compute Units:
 > For the model to run on the ANE, the model must be converted with the `--attention-implementation SPLIT_EINSUM`
 > option.
 > Models converted with `--attention-implementation ORIGINAL` will run on GPU instead of ANE.
 ## Features
 These custom nodes come with a host of features, including:
 - Loading Core ML Unet models
 - Support for ControlNet
 - Support for ANE (Apple Neural Engine)
 - Support for CPU and GPU
 - Support for `mlmodelc` and `mlpackage` files
 - Support for SDXL models
 - Support for LCM models
 - Support for LoRAs
 - SD1.5 -> Core ML conversion
 - SDXL -> Core ML conversion
 - LCM -> Core ML conversion
 > [!NOTE]
 > Please note that using Core ML models can take a bit longer to load initially.
 > For the best experience, I recommend using the compiled models
 > (.mlmodelc files) instead of the .mlpackage files.
 > [!NOTE]  
 > This repository will continue to be updated with more nodes and features over time.
 ## Conversion & Acknowledgements
 The Core ML conversion pipeline in this repository began as an adaptation of
 Apple's [ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion),
 which pioneered running Stable Diffusion on the Apple Neural Engine. The
 implementation has since diverged and no longer depends on that package:
 - UNet conversion runs natively on `diffusers`' `UNet2DConditionModel`.
 - The ANE-friendly attention path (`SPLIT_EINSUM`, `SPLIT_EINSUM_V2`) is
  reimplemented as standalone `diffusers` attention processors.
 - The toolchain tracks current ComfyUI (NumPy 2, Torch 2.7, coremltools 9,
  Python 3.12).
 The goal is to keep iterating on these methods independently and to explore
 support beyond SD1.5.
 > [!IMPORTANT]
 > **Breaking change in 2.0.0.** The converted Core ML UNet now takes
 > `encoder_hidden_states` in the native `diffusers` layout
-> `(batch, tokens, hidden)` instead of the previous
+> `(batch, tokens, hidden)` instead of the previous `(batch, hidden, 1, tokens)`.
-> `(batch, hidden, 1, tokens)`. Core ML models converted with earlier versions
+> Models converted with earlier versions are not compatible and must be
-> are not compatible with 2.0.0 and must be re-converted.
+> re-converted.
-
+
-## Installation
+## Acknowledgements
-
+
-### Using ComfyUI-Manager
+The conversion pipeline began as an adaptation of Apple's
-
+[ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion), which
-The easiest way to install the custom nodes is to use the ComfyUI-Manager. You can find the installation instructions
+pioneered running Stable Diffusion on the Neural Engine. It has since diverged
-[here](https://github.com/ltdrdata/ComfyUI-Manager#installation). Once you've installed the ComfyUI-Manager, you can
+and no longer depends on that package: UNet conversion runs natively on
-install the custom nodes by following these steps:
+`diffusers`' `UNet2DConditionModel`, the ANE attention path (`SPLIT_EINSUM`,
-
+`SPLIT_EINSUM_V2`) is reimplemented as standalone `diffusers` attention
- Open the ComfyUI-Manager by clicking the `Manager` button in the ComfyUI toolbar.
+processors, and the toolchain tracks current ComfyUI (NumPy 2, Torch 2.7+,
- Click the `Install Custom Nodes` button.
+coremltools 9, Python 3.12+). Conversion now lives in the separate
- Search for `Core ML` and click the `Install` button.
+[coreml-diffusion](https://github.com/aszc-dev/coreml-diffusion) package.
 - Restart ComfyUI.
 ### Manual Installation
 1. Clone this repository into the custom_nodes directory of your ComfyUI. If you're not sure how to do this, you can
   download the repository as a zip file and extract it into the same directory.
    ```bash
    cd /path/to/comfyui/custom_nodes
    git clone https://github.com/aszc-dev/ComfyUI-CoreMLSuite.git
    ```
 2. Next, install the required dependencies using pip or another package manager:
    ```bash
    cd /path/to/comfyui/custom_nodes/ComfyUI-CoreMLSuite
    pip install -r requirements.txt
    ```
 ## How to use
 Once you've installed the custom nodes, you can start using them in your ComfyUI workflows.
 To do this, you need to add the nodes to your workflow. You can do this by right-clicking on the workflow canvas and
 selecting the nodes from the list of available nodes (the nodes are in the `Core ML Suite` category).
 You can also double-click the canvas and use the search bar to find the nodes. The list of available nodes is given
 below.
 ### Available Nodes
 #### Core ML UNet Loader (`CoreMLUnetLoader`)
 ![CoreMLUnetLoader](./assets/unet_loader.png?raw=true)
 This node allows you to load a Core ML UNet model and use it in your ComfyUI workflow. Place the converted
 .mlpackage or .mlmodelc file in ComfyUI's `models/unet` directory and use the node to load the model. The output of the
 node is a `coreml_model` object that can be used with the Core ML Sampler.
 - **Inputs**:
    - **model_name**: The name of the model to load. This should be the name of the .mlpackage or .mlmodelc file.
    - **compute_unit**: The hardware on which the model should run. This can be one of the following:
        - `CPU_AND_ANE`: The model will run on both the CPU and ANE. This is the default option. It works best with
          models
          converted with `--attention-implementation SPLIT_EINSUM` or `--attention-implementation SPLIT_EINSUM_V2`.
        - `CPU_AND_GPU`: The model will run on both the CPU and GPU. It works best with models converted with
          `--attention-implementation ORIGINAL`.
        - `CPU_ONLY`: The model will run on the CPU only.
        - `ALL`: The model will run on all available hardware.
 - **Outputs**:
    - **coreml_model**: A Core ML model that can be used with the Core ML Sampler.
 #### Core ML Sampler (`CoreMLSampler`)
 ![CoreMLSampler](./assets/sampler.png?raw=true)
 This node allows you to generate images using a Core ML model. The node takes a Core ML model as input and outputs a
 latent image similar to the latent image output by the KSampler. This means that you can use the
 resulting latent as you normally would in your workflow.
 - **Inputs**:
    - **coreml_model**: The Core ML model to use for sampling. This should be the output of the Core ML UNet Loader.
    - **latent_image** [optional]: The latent image to use for sampling. If provided, should be of the same size as the
      input of the Core ML model. If not provided, the node will create a latent suitable for the Core ML model used.
      Useful in img2img workflows.
    - ... _(the rest of the inputs are the same as the KSampler)_
 - **Outputs**:
    - **LATENT**: The latent image output by the Core ML model. This can be decoded using a VAE Decoder or used as input
      to the next node in your workflow.
 #### Checkpoint Converter
 ![CoreMLConverter](./assets/checkpoint_converter.png?raw=true)
 You can use this node to convert any **SD1.5** based checkpoint to a Core ML model. The converted model is stored in the
 `models/unet` directory and can be used with the `Core ML UNet Loader`. The conversion parameters are encoded in
 the node name, so if the model already exists, the node will not convert it again.
 - **Inputs**:
    - **ckpt_name**: The name of the checkpoint to convert. This should be the name of the checkpoint file stored in the
      `models/checkpoints` directory.
    - **model_version**: Whether the model is based on SD1.5 or SDXL.
    - **height**: The desired height of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
    - **width**: The desired width of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
    - **batch_size**: The batch size of generated images. If you're planning to generate batches of images, you can try
      increasing this value to speed up the generation process. The default is 1.
    - **attention_implementation**: The attention implementation used when converting the model. Choose SPLIT_EINSUM or
      SPLIT_EINSUM_V2 for better ANE support. Choose ORIGINAL for better GPU support.
    - **compute_unit**: The hardware on which the model should run. This is used only when loading the model and doesn't
      affect the conversion process.
    - **controlnet_support**: For the model to support ControlNet, it must be converted with this option set to True.
      The
      default is False.
    - **lora_params** [optional]: Optional LoRA names and weights. If provided, the model will be converted with LoRA(s)
      baked in. More on loading LoRAs below.
 - **Outputs**:
    - **coreml_model**: The converted Core ML model that can be used with Core ML Sampler.
 > [!NOTE]
 > Some models use a custom config .yaml file. If you're using such a model, you'll need to place the config file in the
 > `models/configs` directory. The config file should be named the same as the checkpoint file. For example, if the
 > checkpoint file is named `juggernaut_aftermath.safetensors`, the config file should be
 > named `juggernaut_aftermath.yaml`.
 > The config file will be automatically loaded during conversion.
 > [!NOTE]
 > For now, the converter relies heavilty on the model name to determine the conversion parameters. This means that if
 > you change the model name, the node will convert the model again. Other than that, if you find the name too long or
 > confusing, you can change it to anything you want.
 #### LoRA Loader
 ![LoRALoader](./assets/lora_loader.png?raw=true)
 This node allows you to load LoRAs and bake them into a model. Since this is a workaround (as model weights can't be
 modified
 after conversion), there are a few caveats to keep in mind:
 - The LoRA weights and _strength_model_ parameter are baked into the model. This means that you can't change them
  after conversion. This also means that you need to convert the model again if you want to change the LoRA weights.
 - Loading LoRA affects CLIP, which is not a part of Core ML workflow, so you'll need to load CLIP separately,
  either using `CLIPLoader` or `CheckpointLoaderSimple`. (See [example workflows](#example-workflows) for more details.)
 - After conversion, if you want to load the model using `CoreMLUnetLoader`, you'll need to apply the same LoRAs to
  CLIP manually. (See [example workflows](#example-workflows) for more details.)
 - The LoRA names are encoded in the model name. This means that if you change the name of the LoRA file,
  you'll need to change the model name as well, or the node will convert the model again. (Model strength is not
  encoded, so if you want to change it, you'll need to delete the converted model manually)
 - _strength_clip_ parameter only affects the CLIP model and is not baked into the converted model. This means that
  you can change it after conversion.
 - **Inputs**:
    - **lora_name**: The name of the LoRA to load.
    - **strength_model**: The strength of the LoRA model.
    - **strength_clip**: The strength of the LoRA CLIP.
    - **lora_params** [optional]: Optional output from other LoRA Loaders.
    - **clip**: The CLIP model to use with the LoRA. This can be either output of the
      `CLIPLoader`/`CheckpointLoaderSimple` or other LoRA Loaders.
 - **Outputs**:
    - **lora_params**: The LoRA parameters that can be passed to the Core ML Converter or other LoRA Loaders.
    - **CLIP**: The CLIP model with LoRA applied.
 #### LCM Converter
 ![LCMConverter](./assets/lcm_converter.png?raw=true)
 This node converts [SimianLuo/LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7) model to Core
 ML. The converted model is stored in the `models/unet` directory and can be used with the Core ML UNet Loader. The
 conversion parameteres are encoded in the node name, so if the model already exists, the node will not convert it again.
 - **Inputs**:
    - **height**: The desired height of the image generated by the model. The default is 512. Must be a multiple of 8.
    - **width**: The desired width of the image generated by the model. The default is 512. Must be a multiple of 8.
    - **batch_size**: The batch size of generated images. If you're planning to generate batches of images, you can try
      increasing this value to speed up the generation process. The default is 1.
    - **compute_unit**: The hardware on which the model should run. This is used only when loading the model and
      doesn't affect the conversion process.
    - **controlnet_support**: For the model to support ControlNet, it must be converted with this option set to True.
      The default is False.
 > [!NOTE]
 > The conversion process can take a while, so please be patient.
 > [!NOTE]
 > When using the LCM model with Core ML Sampler, please set _sampler_name_ to `lcm` and _scheduler_ to `sgm_uniform`.
 #### Core ML Adapter (Experimental) (`CoreMLModelAdapter`)
 ![CoreMLModelAdapter](./assets/adapter.png?raw=true)
 This node allows you to use a Core ML as a standard ComfyUI model. This is an experimental node and may not work with
 all models and nodes. Please use with caution and pay attention to the expected inputs of the model.
 - **Input**:
    - **coreml_model**: The Core ML model to use as a ComfyUI model.
 - **Output**:
    - **MODEL**: The Core ML model wrapped in a ComfyUI model.
 > [!NOTE]
 > While this approach allows you to use Core ML models with many ComfyUI nodes (both standard and custom), the
 > expected inputs of the model will not be checked, which may cause errors. Please make sure to use a model compatible
 > with the expected parameters.
 ### Example Workflows
 > [!NOTE]
 > The models used are just an example. Feel free to experiment with different models and see what works best for you.
 #### Basic txt2img with Core ML UNet loader
 This is a basic txt2img workflow that uses the Core ML UNet loader to load a model. The CLIP and VAE models
 are loaded using the standard ComfyUI nodes. In the first example, the text encoder (CLIP) and VAE models are loaded
 separately. In the second example, the text encoder and VAE models are loaded from the checkpoint file. Note that you
 can use any CLIP or VAE model as long as it's compatible with Stable Diffusion v1.5.
 1. **Loading text encoder (CLIP) and VAE models separately**
    - This workflow uses CLIP and VAE models available
      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/text_encoder/model.safetensors) and
      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/vae/diffusion_pytorch_model.safetensors).
      Once downloaded, place the models in the`models/clip` and `models/vae` directories respectively.
    - The Core ML UNet model is available
      [here](https://huggingface.co/coreml-community/coreml-stable-diffusion-v1-5_cn/blob/main/split_einsum/stable-diffusion-_v1-5_split-einsum_cn.zip).
      Once downloaded, place the model in the `models/unet` directory.  
      ![coreml-unet+clip+vae](./assets/unet+sampler+clip+vae.png?raw=true)
 2. **Loading text encoder (CLIP) and VAE models from checkpoint file**
    - This workflow loads the CLIP and VAE models from the checkpoint file available
      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors).
      Once downloaded, place the model in the`models/checkpoints` directory.
    - The Core ML UNet model is available
      [here](https://huggingface.co/coreml-community/coreml-stable-diffusion-v1-5_cn/blob/main/split_einsum/stable-diffusion-_v1-5_split-einsum_cn.zip).
      Once downloaded, place the model in the `models/unet` directory.  
      ![coreml-unet+checkpoint](./assets/unet+sampler+checkpoint.png?raw=true)
 #### ControlNet with Core ML UNet loader
 This workflow uses the Core ML UNet loader to load a Core ML UNet model that supports ControlNet. The ControlNet is
 being loaded using the standard ComfyUI nodes. Please refer to
 the [basic txt2img workflow](#basic-txt2img-with-core-ml-unet-loader) for more details on how to load the CLIP and VAE
 models.
 The ControlNet model used in this workflow is available
 [here](https://huggingface.co/lllyasviel/control_v11p_sd15_scribble/blob/main/diffusion_pytorch_model.fp16.safetensors).
 Once downloaded, place the model in the `models/controlnet` directory.
 ![coreml-unet+controlnet](./assets/unet+sampler+controlnet.png?raw=true)
 #### Checkpoint conversion
 This workflow uses the Checkpoint Converter to convert the checkpoint file. See
 [Checkpoint Converter](#checkpoint-converter) description for more details.
 ![checkpoint-converter](./assets/basic_conversion.png?raw=true)
 #### Checkpoint conversion with LoRA
 This workflow uses the Checkpoint Converter to convert the checkpoint file with LoRA. See
 [LoRA Loader](#lora-loader) description to read more about the caveats of using LoRA.
 ![checkpoint-converter+lora](./assets/conversion+lora.png?raw=true)
 #### LCM LoRA conversion
 Please note that you can use multiple LoRAs with the same model. To do this, you'll need to use multiple LoRA Loaders.
 > [!IMPORTANT]
 > In this example, the model is passed through the adapter and `ModelSamplingDiscrete` nodes to a standard ComfyUI's
 > KSampler (not Core ML Sampler). ModelSamplingDiscrete needs to be used to sample models with LCM LoRAs properly.
 ![multiple-loras](./assets/conversion+lcm_lora.png?raw=true)
 #### Loader with LoRAs
 This workflow uses the Core ML UNet Loader to load a model with LoRAs. The CLIP must be loaded separately and passed
 through the same LoRA nodes as during conversion. See [LoRA Loader](#lora-loader) description to read more about the
 caveats of using LoRA. Since _lora_name_ and _strength_model_ are baked into the model, it is not necessary to pass
 them as inputs to the loader.
 > [!IMPORTANT]
 > In this example, the model is passed through the adapter and `ModelSamplingDiscrete` nodes to a standard ComfyUI's
 > KSampler (not Core ML Sampler). ModelSamplingDiscrete needs to be used to sample models with LCM LoRAs properly.
 ![loader+lora](./assets/loader+lcm_lora.png?raw=true)
 #### LCM conversion with ControlNet
 This workflow uses LCM converter to
 convert [SimianLuo/LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7)
 model to Core ML. The converted model can then be used with or without ControlNet to generate images.
 ![lcm+controlnet](./assets/lcm+controlnet.png?raw=true)
 #### SDXL Base + Refiner conversion
 This is a basic workflow for SDXL. You add LoRAs and ControlNets the same way as in the previous examples.
 You can also skip the refiner step.
 The models used in this workflow are available at the following links:
 - [Base model + text_encoder (clip) + text_encoder_2 (clip2)](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
 - [Refiner model](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)
 - [VAE](https://huggingface.co/stabilityai/sdxl-vae)
 > [!IMPORTANT]
 > SDXL on ANE is not supported. If loading of the model gets stuck, please try using CPU_AND_GPU or CPU_ONLY.  
 > For best results, use ORIGINAL attention implementation.
 ![sdxl](./assets/sdxl_conversion.png?raw=true)
 ## Quantization (opt-in)
 The `Core ML Converter` and `Core ML LCM Converter` nodes accept an
 optional `quantize_nbits` dropdown that runs k-means weight palettization
 (`coremltools.optimize.coreml.palettize_weights`) on the UNet before save.
 Values: `none` (default — no quantization, identical to unquantized
 behavior and filenames), `8`, `6`, `4`. The number is appended to the
 .mlpackage stem as `_q<bits>` so quantized and unquantized variants
 coexist on disk and in cache.
 ### SD1.5 1×512×512 SPLIT_EINSUM tradeoffs (M2 Pro, ANE)
 Measured with 20 UNet forward passes at a fixed seed for the PSNR
 comparison:
 | nbits | size (MB) | size vs none | fwd median (ms) | PSNR vs `none` (dB) |
 |---|---:|---:|---:|---:|
 | none | 1641 | 1.000 | 197.1 | — |
 | 8    |  822 | 0.501 | 186.6 | 53.5 |
 | 6    |  617 | 0.376 | 183.0 | 40.2 |
 | 4    |  412 | 0.251 | 179.8 | 27.5 |
 PSNR here is computed on the raw `noise_pred` output of a single UNet
 forward at a fixed seed, not on the final decoded image — it isolates
 the quantization-induced drift from sampler / VAE noise. Final-image
 PSNR is comfortably higher (the sampler averages over 20 steps).
 ### Recommended settings per chip / RAM
 - **8 GB RAM (M1 base, M2 base):** `nbits=4`. ~4× smaller model, still
  loads, PSNR 27 dB is visually identical at SD1.5 sizes.
 - **16 GB RAM (M1/M2/M3 Pro):** `nbits=6` is the sweet spot — ~2.7×
  smaller, PSNR 40 dB, no perceptible quality drop.
 - **32 GB+ RAM (Max / Ultra):** `nbits=8` if you want the safety
  margin, `none` if you want bit-identical output for golden testing.
 The default stays `none` so existing workflows produce byte-for-byte
 identical output.
 ## Limitations
 - Core ML models are fixed in terms of their inputs and outputs.
  This means you'll need to use latent images of the same size as the input of the model (512x512 is the default for
  SD1.5).
  However, you can re-convert the model to a different input size using the
  conversion nodes in this suite (set the desired width and height).
 - SD2.1 models are not supported.
 [^1]:
 Unless [EnumeratedShapes](https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html#select-from-predetermined-shapes)
 is used during conversion. Needs more testing.
 ## Support
-I'm here to help! If you have any questions or suggestions, don't hesitate to open an issue and I'll do my best
+Questions or suggestions? Open an
-to assist you.
+[issue](https://github.com/aszc-dev/ComfyUI-CoreMLSuite/issues).
@@ -11,9 +11,6 @@ from coreml_suite.nodes import (
    CoreMLConverter,
    COREML_LOAD_LORA,
 )
 from coreml_suite.lcm import (
    COREML_CONVERT_LCM,
 )
 NODE_CLASS_MAPPINGS = {
    "CoreMLUNetLoader": CoreMLLoaderUNet,
@@ -22,7 +19,6 @@ NODE_CLASS_MAPPINGS = {
    "CoreMLModelAdapter": CoreMLModelAdapter,
    "Core ML LoRA Loader": COREML_LOAD_LORA,
    "Core ML Converter": CoreMLConverter,
    "Core ML LCM Converter": COREML_CONVERT_LCM,
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
    "CoreMLUNetLoader": "Load Core ML UNet",
@@ -31,5 +27,4 @@ NODE_DISPLAY_NAME_MAPPINGS = {
    "CoreMLModelAdapter": "Core ML Adapter (Experimental)",
    "Core ML LoRA Loader": "Load LoRA to use with Core ML",
    "Core ML Converter": "Convert Checkpoint to Core ML",
    "Core ML LCM Converter": "Convert LCM to Core ML",
 }
@@ -1,5 +0,0 @@
 ATTENTION_IMPLEMENTATIONS = (
    "SPLIT_EINSUM",
    "SPLIT_EINSUM_V2",
    "ORIGINAL",
 )
@@ -4,7 +4,7 @@ from comfy import supported_models_base
 from comfy import latent_formats
 from comfy.model_detection import convert_config
-from coreml_suite.model_version import ModelVersion
+from coreml_diffusion import ModelVersion
 config_map = {
@@ -1,9 +0,0 @@
 """Core ML conversion helpers.
 The conversion approach originates from Apple's ml-stable-diffusion
 (https://github.com/apple/ml-stable-diffusion). This implementation has since
 diverged: it runs natively on diffusers' UNet2DConditionModel with its own
 SPLIT_EINSUM / SPLIT_EINSUM_V2 attention processors and no longer depends on
 that package. The intent is to keep iterating on these methods independently
 while tracking current tooling.
 """
@@ -1,239 +0,0 @@
 import logging
 import torch
 logger = logging.getLogger(__name__)
 CHUNK_SIZE = 512
 def apply_attention_implementation(unet, attention_implementation):
    if attention_implementation == "ORIGINAL":
        return unet
    if attention_implementation == "SPLIT_EINSUM":
        unet.set_attn_processor(SplitEinsumAttnProcessor())
        return unet
    if attention_implementation == "SPLIT_EINSUM_V2":
        unet.set_attn_processor(SplitEinsumV2AttnProcessor())
        return unet
    raise ValueError(f"Unsupported attention implementation: {attention_implementation}")
 class SplitEinsumAttnProcessor:
    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        *args,
        **kwargs,
    ):
        return _attention_forward(
            attn,
            hidden_states,
            encoder_hidden_states,
            attention_mask,
            temb,
            split_einsum,
        )
 class SplitEinsumV2AttnProcessor:
    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        *args,
        **kwargs,
    ):
        return _attention_forward(
            attn,
            hidden_states,
            encoder_hidden_states,
            attention_mask,
            temb,
            split_einsum_v2,
        )
 def _attention_forward(
    attn,
    hidden_states,
    encoder_hidden_states,
    attention_mask,
    temb,
    attention_fn,
 ):
    residual = hidden_states
    if attn.spatial_norm is not None:
        hidden_states = attn.spatial_norm(hidden_states, temb)
    input_ndim = hidden_states.ndim
    if input_ndim == 4:
        batch_size, channel, height, width = hidden_states.shape
        hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
    else:
        batch_size, _, channel = hidden_states.shape
        height = None
        width = None
    batch_size, key_sequence_length, _ = (
        hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
    )
    if attention_mask is not None:
        attention_mask = attn.prepare_attention_mask(
            attention_mask,
            key_sequence_length,
            batch_size,
        )
        attention_mask = _prepare_split_einsum_mask(
            attention_mask,
            batch_size,
            attn.heads,
            key_sequence_length,
        )
    if attn.group_norm is not None:
        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
    query = attn.to_q(hidden_states)
    if encoder_hidden_states is None:
        encoder_hidden_states = hidden_states
    elif attn.norm_cross:
        encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
    key = attn.to_k(encoder_hidden_states)
    value = attn.to_v(encoder_hidden_states)
    batch_size = query.shape[0]
    dim_head = attn.inner_kv_dim // attn.heads
    query = _linear_projection_to_bchw(query)
    key = _linear_projection_to_bchw(key)
    value = _linear_projection_to_bchw(value)
    hidden_states = attention_fn(
        query,
        key,
        value,
        attention_mask,
        attn.heads,
        dim_head,
    )
    hidden_states = hidden_states.squeeze(2).transpose(1, 2)
    hidden_states = hidden_states.reshape(batch_size, -1, attn.inner_dim)
    hidden_states = attn.to_out[0](hidden_states)
    hidden_states = attn.to_out[1](hidden_states)
    if input_ndim == 4:
        hidden_states = hidden_states.transpose(-1, -2).reshape(
            batch_size,
            channel,
            height,
            width,
        )
    if attn.residual_connection:
        hidden_states = hidden_states + residual
    hidden_states = hidden_states / attn.rescale_output_factor
    return hidden_states
 def split_einsum(q, k, v, mask, heads, dim_head):
    q_heads = _split_heads(q, heads, dim_head)
    k = k.transpose(1, 3)
    k_heads = [
        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
        for head_idx in range(heads)
    ]
    v_heads = _split_heads(v, heads, dim_head)
    weights = [
        torch.einsum("bchq,bkhc->bkhq", query, key) * (dim_head**-0.5)
        for query, key in zip(q_heads, k_heads)
    ]
    if mask is not None:
        weights = [weight + mask for weight in weights]
    weights = [weight.softmax(dim=1) for weight in weights]
    outputs = [
        torch.einsum("bkhq,bchk->bchq", weight, value)
        for weight, value in zip(weights, v_heads)
    ]
    return torch.cat(outputs, dim=1)
 def split_einsum_v2(q, k, v, mask, heads, dim_head):
    query_length = q.size(3)
    num_chunks = query_length // CHUNK_SIZE
    if num_chunks == 0:
        logger.info(
            "SPLIT_EINSUM_V2 query sequence is shorter than %s; using SPLIT_EINSUM.",
            CHUNK_SIZE,
        )
        return split_einsum(q, k, v, mask, heads, dim_head)
    q_heads = _split_heads(q, heads, dim_head)
    q_chunks = [
        [
            head[..., chunk_idx * CHUNK_SIZE : (chunk_idx + 1) * CHUNK_SIZE]
            for chunk_idx in range(num_chunks)
        ]
        for head in q_heads
    ]
    k = k.transpose(1, 3)
    k_heads = [
        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
        for head_idx in range(heads)
    ]
    v_heads = _split_heads(v, heads, dim_head)
    head_outputs = []
    for query_chunks, key, value in zip(q_chunks, k_heads, v_heads):
        chunk_outputs = []
        for query_chunk in query_chunks:
            weights = torch.einsum("bchq,bkhc->bkhq", query_chunk, key)
            weights = weights * (dim_head**-0.5)
            if mask is not None:
                weights = weights + mask
            weights = weights.softmax(dim=1)
            chunk_outputs.append(torch.einsum("bkhq,bchk->bchq", weights, value))
        head_outputs.append(torch.cat(chunk_outputs, dim=3))
    return torch.cat(head_outputs, dim=1)
 def _split_heads(x, heads, dim_head):
    return [
        x[:, head_idx * dim_head : (head_idx + 1) * dim_head, :, :]
        for head_idx in range(heads)
    ]
 def _linear_projection_to_bchw(x):
    return x.transpose(1, 2).unsqueeze(2)
 def _prepare_split_einsum_mask(mask, batch_size, heads, key_sequence_length):
    if mask.ndim == 2:
        mask = mask[:, None, :]
    if mask.shape[0] == batch_size * heads:
        mask = mask.reshape(batch_size, heads, -1, key_sequence_length)
        mask = mask[:, 0]
    if mask.ndim == 3:
        mask = mask[:, :, None, None]
    return mask
@@ -1,20 +0,0 @@
 def conv2d_output_shape(height, width, conv):
    """Return the spatial output shape for a torch.nn.Conv2d-like module."""
    kernel_h, kernel_w = _pair(conv.kernel_size)
    stride_h, stride_w = _pair(conv.stride)
    pad_h, pad_w = _pair(conv.padding)
    dilation_h, dilation_w = _pair(conv.dilation)
    out_h = _conv_output_dim(height, kernel_h, stride_h, pad_h, dilation_h)
    out_w = _conv_output_dim(width, kernel_w, stride_w, pad_w, dilation_w)
    return out_h, out_w
 def _conv_output_dim(size, kernel, stride, padding, dilation):
    return ((size + (2 * padding) - (dilation * (kernel - 1)) - 1) // stride) + 1
 def _pair(value):
    if isinstance(value, tuple):
        return value
    return value, value
@@ -1,61 +0,0 @@
 from types import MethodType
 from diffusers.models.transformers.transformer_2d import Transformer2DModel
 def prepare_unet_for_coreml_trace(unet):
    for module in unet.modules():
        if isinstance(module, Transformer2DModel):
            module._operate_on_continuous_inputs = MethodType(
                _operate_on_continuous_inputs,
                module,
            )
            module._get_output_for_continuous_inputs = MethodType(
                _get_output_for_continuous_inputs,
                module,
            )
    return unet
 def _operate_on_continuous_inputs(self, hidden_states):
    hidden_states = self.norm(hidden_states)
    if not self.use_linear_projection:
        hidden_states = self.proj_in(hidden_states)
        inner_dim = self.inner_dim
        hidden_states = hidden_states.flatten(2).transpose(1, 2)
    else:
        inner_dim = hidden_states.shape[1]
        hidden_states = hidden_states.flatten(2).transpose(1, 2)
        hidden_states = self.proj_in(hidden_states)
    return hidden_states, inner_dim
 def _get_output_for_continuous_inputs(
    self,
    hidden_states,
    residual,
    batch_size,
    height,
    width,
    inner_dim,
 ):
    if not self.use_linear_projection:
        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size,
            inner_dim,
            height,
            width,
        )
        hidden_states = self.proj_out(hidden_states)
    else:
        hidden_states = self.proj_out(hidden_states)
        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size,
            inner_dim,
            height,
            width,
        )
    return hidden_states + residual
@@ -1,54 +0,0 @@
 import torch
 class CoreMLUNetWrapper(torch.nn.Module):
    """Adapt diffusers UNet inputs to CoreMLSuite's stable Core ML contract."""
    def __init__(self, unet, model_version):
        super().__init__()
        self.unet = unet
        self.model_version = model_version
    def forward(self, sample, timestep, encoder_hidden_states, *extra_inputs):
        input_index = 0
        timestep_cond = None
        if self._is_lcm:
            timestep_cond = extra_inputs[input_index]
            input_index += 1
        added_cond_kwargs = None
        if self._is_sdxl:
            time_ids = extra_inputs[input_index]
            text_embeds = extra_inputs[input_index + 1]
            input_index += 2
            added_cond_kwargs = {
                "time_ids": time_ids,
                "text_embeds": text_embeds,
            }
        additional_residuals = extra_inputs[input_index:]
        down_residuals = None
        mid_residual = None
        if additional_residuals:
            down_residuals = tuple(additional_residuals[:-1])
            mid_residual = additional_residuals[-1]
        outputs = self.unet(
            sample,
            timestep,
            encoder_hidden_states=encoder_hidden_states,
            timestep_cond=timestep_cond,
            added_cond_kwargs=added_cond_kwargs,
            down_block_additional_residuals=down_residuals,
            mid_block_additional_residual=mid_residual,
            return_dict=False,
        )
        return outputs[0]
    @property
    def _is_lcm(self):
        return self.model_version.name == "LCM"
    @property
    def _is_sdxl(self):
        return self.model_version.name in {"SDXL", "SDXL_REFINER"}
@@ -1,322 +0,0 @@
 import gc
 import os
 import time
 import coremltools as ct
 import numpy as np
 import torch
 from diffusers import UNet2DConditionModel
 from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
 from coreml_suite.conversion.attention import apply_attention_implementation
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.trace import prepare_unet_for_coreml_trace
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 from coreml_suite.logger import logger
 from coreml_suite.model_version import ModelVersion
 DEFAULT_TRACE_TIMESTEP = 999.0
 TEXT_TOKEN_SEQUENCE_LENGTH = 77
 def get_unet(model_version: ModelVersion, ref_unet, attention_implementation):
    ref_unet = prepare_unet_for_coreml_trace(ref_unet)
    unet = apply_attention_implementation(
        ref_unet.eval(),
        attention_implementation,
    )
    return CoreMLUNetWrapper(unet, model_version)
 def get_encoder_hidden_states_shape(ref_unet, batch_size):
    encoder_hidden_states_shape = (
        batch_size,
        TEXT_TOKEN_SEQUENCE_LENGTH,
        ref_unet.config.cross_attention_dim,
    )
    return encoder_hidden_states_shape
 def get_coreml_inputs(sample_inputs):
    coreml_sample_unet_inputs = {
        k: v.numpy().astype(np.float16) for k, v in sample_inputs.items()
    }
    return [
        ct.TensorType(
            name=k,
            shape=v.shape,
            dtype=v.numpy().dtype if isinstance(v, torch.Tensor) else v.dtype,
        )
        for k, v in coreml_sample_unet_inputs.items()
    ]
 def load_coreml_model(out_path):
    logger.info(f"Loading model from {out_path}")
    start = time.time()
    coreml_model = ct.models.MLModel(out_path)
    logger.info(f"Loading {out_path} took {time.time() - start:.1f} seconds")
    return coreml_model
 def convert_to_coreml(
    submodule_name, torchscript_module, sample_inputs, output_names, out_path
 ):
    if os.path.exists(out_path):
        logger.info(f"Skipping export because {out_path} already exists")
        coreml_model = load_coreml_model(out_path)
    else:
        logger.info(f"Converting {submodule_name} to CoreML..")
        coreml_model = ct.convert(
            torchscript_module,
            convert_to="mlprogram",
            minimum_deployment_target=ct.target.macOS13,
            inputs=sample_inputs,
            outputs=[
                ct.TensorType(name=name, dtype=np.float32) for name in output_names
            ],
            skip_model_load=True,
        )
        del torchscript_module
        gc.collect()
    return coreml_model
 def get_out_path(submodule_name, model_name):
    from folder_paths import get_folder_paths
    fname = f"{model_name}_{submodule_name}.mlpackage"
    unet_path = get_folder_paths(submodule_name)[0]
    out_path = os.path.join(unet_path, fname)
    return out_path
 def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape):
    sample_unet_inputs = dict(
        [
            ("sample", torch.rand(*sample_shape)),
            (
                "timestep",
                torch.tensor([DEFAULT_TRACE_TIMESTEP] * batch_size).to(torch.float32),
            ),
            ("encoder_hidden_states", torch.rand(*encoder_hidden_states_shape)),
        ]
    )
    return sample_unet_inputs
 def lcm_inputs(sample_unet_inputs):
    batch_size = sample_unet_inputs["sample"].shape[0]
    return {"timestep_cond": torch.randn(batch_size, 256).to(torch.float32)}
 def sdxl_inputs(sample_unet_inputs, ref_unet, model_version):
    sample_shape = sample_unet_inputs["sample"].shape
    batch_size = sample_shape[0]
    h = sample_shape[2] * 8
    w = sample_shape[3] * 8
    original_size = (h, w)
    crops_coords_top_left = (0, 0)
    is_refiner = model_version == ModelVersion.SDXL_REFINER
    if is_refiner:
        aesthetic_score = (6.0,)
        time_ids_list = list(original_size + crops_coords_top_left + aesthetic_score)
    else:
        target_size = (h, w)
        time_ids_list = list(original_size + crops_coords_top_left + target_size)
    time_ids = torch.tensor(time_ids_list).repeat(batch_size, 1).to(torch.int64)
    text_embeds_shape = (batch_size, get_sdxl_text_embeds_dim(ref_unet, len(time_ids_list)))
    return {
        "time_ids": time_ids,
        "text_embeds": torch.randn(*text_embeds_shape).to(torch.float32),
    }
 def get_sdxl_text_embeds_dim(ref_unet, time_ids_dim):
    projection_dim = ref_unet.config.projection_class_embeddings_input_dim
    time_embed_dim = ref_unet.config.addition_time_embed_dim
    return projection_dim - (time_ids_dim * time_embed_dim)
 def get_inputs_spec(inputs):
    inputs_spec = {k: (v.shape, v.dtype) for k, v in inputs.items()}
    return inputs_spec
 def add_cnet_support(sample_shape, reference_unet):
    additional_residuals_shapes = []
    batch_size = sample_shape[0]
    h, w = sample_shape[2:]
    # conv_in
    out_h, out_w = conv2d_output_shape(
        h,
        w,
        reference_unet.conv_in,
    )
    additional_residuals_shapes.append(
        (batch_size, reference_unet.conv_in.out_channels, out_h, out_w)
    )
    # down_blocks
    for down_block in reference_unet.down_blocks:
        additional_residuals_shapes += [
            (batch_size, resnet.out_channels, out_h, out_w)
            for resnet in down_block.resnets
        ]
        if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
            for downsampler in down_block.downsamplers:
                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
            additional_residuals_shapes.append(
                (
                    batch_size,
                    down_block.downsamplers[-1].conv.out_channels,
                    out_h,
                    out_w,
                )
            )
    # mid_block
    additional_residuals_shapes.append(
        (batch_size, reference_unet.mid_block.resnets[-1].out_channels, out_h, out_w)
    )
    additional_inputs = {}
    for i, shape in enumerate(additional_residuals_shapes):
        sample_residual_input = torch.rand(*shape)
        additional_inputs[f"additional_residual_{i}"] = sample_residual_input
    return additional_inputs
 def convert_unet(
    ref_unet,
    model_version: ModelVersion,
    unet_out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    attention_implementation: str = ATTENTION_IMPLEMENTATIONS[0],
    quantize_nbits: str = "none",
 ):
    coreml_unet = get_unet(model_version, ref_unet, attention_implementation)
    sample_shape = (
        batch_size,  # B
        ref_unet.config.in_channels,  # C
        sample_size[0],  # H
        sample_size[1],  # W
    )
    encoder_hidden_states_shape = get_encoder_hidden_states_shape(ref_unet, batch_size)
    sample_inputs = get_sample_input(
        batch_size, encoder_hidden_states_shape, sample_shape
    )
    if model_version == ModelVersion.LCM:
        sample_inputs |= lcm_inputs(sample_inputs)
    if model_version in {ModelVersion.SDXL, ModelVersion.SDXL_REFINER}:
        sample_inputs |= sdxl_inputs(sample_inputs, ref_unet, model_version)
    if controlnet_support:
        sample_inputs |= add_cnet_support(sample_shape, ref_unet)
    sample_inputs_spec = get_inputs_spec(sample_inputs)
    logger.info(f"Sample UNet inputs spec: {sample_inputs_spec}")
    logger.info("JIT tracing..")
    traced_unet = torch.jit.trace(
        coreml_unet, example_inputs=list(sample_inputs.values())
    )
    logger.info("Done.")
    coreml_sample_inputs = get_coreml_inputs(sample_inputs)
    coreml_unet = convert_to_coreml(
        "unet", traced_unet, coreml_sample_inputs, ["noise_pred"], unet_out_path
    )
    del traced_unet
    gc.collect()
    if quantize_nbits != "none":
        # Opt-in k-means weight palettization. The default path
        # (quantize_nbits="none") leaves the traced UNet untouched.
        from coremltools.optimize.coreml import (
            OpPalettizerConfig,
            OptimizationConfig,
            palettize_weights,
        )
        nbits = int(quantize_nbits)
        logger.info(f"Palettizing UNet weights to {nbits}-bit (kmeans)..")
        t0 = time.time()
        cfg = OptimizationConfig(
            global_config=OpPalettizerConfig(mode="kmeans", nbits=nbits)
        )
        coreml_unet = palettize_weights(coreml_unet, config=cfg)
        logger.info(f"Palettization took {time.time() - t0:.1f}s")
    coreml_unet.save(unet_out_path)
    logger.info(f"Saved unet into {unet_out_path}")
 def convert(
    ckpt_path: str,
    model_version: ModelVersion,
    unet_out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    lora_weights: list[tuple[str | os.PathLike, float]] = None,
    attn_impl: str = ATTENTION_IMPLEMENTATIONS[0],
    config_path: str = None,
    quantize_nbits: str = "none",
 ):
    if os.path.exists(unet_out_path):
        logger.info(f"Found existing model at {unet_out_path}! Skipping..")
        return
    if attn_impl not in ATTENTION_IMPLEMENTATIONS:
        raise ValueError(
            f"Unsupported attention implementation {attn_impl!r}. "
            f"Expected one of {ATTENTION_IMPLEMENTATIONS}."
        )
    ref_unet = load_unet(ckpt_path, config_path)
    for i, lora_weight in enumerate(lora_weights or []):
        lora_path, strength = lora_weight
        adapter_name = f"lora_{i}"
        ref_unet.load_lora_adapter(lora_path, adapter_name=adapter_name)
        ref_unet.set_adapters([adapter_name], weights=[strength])
        ref_unet.fuse_lora()
    convert_unet(
        ref_unet,
        model_version,
        unet_out_path,
        batch_size,
        sample_size,
        controlnet_support,
        attention_implementation=attn_impl,
        quantize_nbits=quantize_nbits,
    )
 def load_unet(ckpt_path, config_path):
    return UNet2DConditionModel.from_single_file(
        ckpt_path,
        original_config=config_path,
    )
@@ -1,68 +0,0 @@
 """Pure out_name composition for the Core ML UNet artifact.
 Extracted from CoreMLConverter.convert so the filename contract
 can be tested + reused without instantiating the node. The string is the
 cache key: every workflow that references a converted .mlpackage depends
 on it staying byte-for-byte identical.
 """
 from typing import Iterable, Tuple
 ATTN_SUFFIX = {
    "SPLIT_EINSUM": "se",
    "SPLIT_EINSUM_V2": "se2",
    "ORIGINAL": "orig",
 }
 # Palettization bits. "none" = no quantization (default; keeps the
 # unquantized filename intact so existing workflows still resolve their
 # cached .mlpackage). Numeric values append a `_q<bits>` suffix.
 QUANT_NBITS_VALUES = ("none", "8", "6", "4")
 def compose_out_name(
    *,
    ckpt_name: str,
    batch_size: int,
    width: int,
    height: int,
    controlnet_support: bool,
    attention_implementation: str,
    lora_names: Iterable[str] = (),
    quantize_nbits: str = "none",
 ) -> str:
    """Build the .mlpackage stem from convert() parameters.
    Locked behaviour (characterization tests):
      - first '.' in ckpt_name wins (`a.b.c.safetensors` -> `a`)
      - spaces collapse to underscores
      - LoRA names are taken stem-only, sorted, joined with '_' and
        prefixed with '_' when present (caller is expected to pass a
        sorted list; we sort defensively)
      - controlnet adds `_cn`
      - attn suffix is `_se` | `_se2` | `_orig`
    Quantization:
      - quantize_nbits "none" (default) appends nothing — existing
        unquantized .mlpackages keep the old filename
      - "4" / "6" / "8" appends `_q<bits>` after the attn suffix
    """
    if quantize_nbits not in QUANT_NBITS_VALUES:
        raise ValueError(
            f"quantize_nbits={quantize_nbits!r} not in {QUANT_NBITS_VALUES}"
        )
    stem = ckpt_name.split(".")[0]
    sorted_names = sorted(lora_names)
    lora_str = "_" + "_".join(name.split(".")[0] for name in sorted_names) if sorted_names else ""
    cn_suffix = "_cn" if controlnet_support else ""
    attn_suffix = "_" + ATTN_SUFFIX[attention_implementation]
    quant_suffix = f"_q{quantize_nbits}" if quantize_nbits != "none" else ""
    out_name = (
        f"{stem}{lora_str}_{batch_size}x{width}x{height}"
        f"{cn_suffix}{attn_suffix}{quant_suffix}"
    )
    return out_name.replace(" ", "_")
 def lora_names_from_params(lora_params: Iterable[Tuple[str, float]]) -> list[str]:
    """Mirror the sort applied inside CoreMLConverter.convert."""
    return [name for name, _ in sorted(lora_params, key=lambda pair: pair[0])]
@@ -1,3 +1,8 @@
-from .nodes import COREML_CONVERT_LCM
+"""LCM runtime support (sampler-side).
-__all__ = ["COREML_CONVERT_LCM"]
+The dedicated LCM converter node was removed once the standard ``CoreMLConverter``
 gained model-version auto-detection (full-distill LCM is detected from the
 checkpoint). What remains here is runtime sampling support — ``utils`` patches the
 model sampling and supplies the guidance embedding when a converted UNet exposes
 ``timestep_cond``.
 """
@@ -1,259 +0,0 @@
 import os
 import logging
 import time
 import gc
 import numpy as np
 import torch
 from diffusers import UNet2DConditionModel, LCMScheduler
 from diffusers.loaders import LoraLoaderMixin
 from coreml_suite.conversion.attention import apply_attention_implementation
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 from coreml_suite.model_version import ModelVersion
 import coremltools as ct
 logging.basicConfig()
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 MODEL_VERSION = "SimianLuo/LCM_Dreamshaper_v7"
 MODEL_NAME = MODEL_VERSION.split("/")[-1] + "_4k"
 TEXT_TOKEN_SEQUENCE_LENGTH = 77
 def get_unets():
    ref_unet = UNet2DConditionModel.from_pretrained(
        MODEL_VERSION,
        subfolder="unet",
        device_map=None,
        low_cpu_mem_usage=False,
    )
    cml_unet = CoreMLUNetWrapper(
        apply_attention_implementation(ref_unet.eval(), "SPLIT_EINSUM"),
        ModelVersion.LCM,
    )
    return cml_unet, ref_unet
 def get_encoder_hidden_states_shape(unet_config, batch_size):
    encoder_hidden_states_shape = (
        batch_size,
        TEXT_TOKEN_SEQUENCE_LENGTH,
        unet_config.cross_attention_dim,
    )
    return encoder_hidden_states_shape
 def get_scheduler():
    from comfy.model_management import get_torch_device
    scheduler = LCMScheduler.from_pretrained(MODEL_VERSION, subfolder="scheduler")
    scheduler.set_timesteps(50, get_torch_device(), 50)
    return scheduler
 def get_coreml_inputs(sample_inputs):
    coreml_sample_unet_inputs = {
        k: v.numpy().astype(np.float16) for k, v in sample_inputs.items()
    }
    return [
        ct.TensorType(
            name=k,
            shape=v.shape,
            dtype=v.numpy().dtype if isinstance(v, torch.Tensor) else v.dtype,
        )
        for k, v in coreml_sample_unet_inputs.items()
    ]
 def load_coreml_model(out_path):
    logger.info(f"Loading model from {out_path}")
    start = time.time()
    coreml_model = ct.models.MLModel(out_path)
    logger.info(f"Loading {out_path} took {time.time() - start:.1f} seconds")
    return coreml_model
 def convert_to_coreml(
    submodule_name, torchscript_module, sample_inputs, output_names, out_path
 ):
    if os.path.exists(out_path):
        logger.info(f"Skipping export because {out_path} already exists")
        coreml_model = load_coreml_model(out_path)
    else:
        logger.info(f"Converting {submodule_name} to CoreML..")
        coreml_model = ct.convert(
            torchscript_module,
            convert_to="mlprogram",
            minimum_deployment_target=ct.target.macOS13,
            inputs=sample_inputs,
            outputs=[
                ct.TensorType(name=name, dtype=np.float32) for name in output_names
            ],
            skip_model_load=True,
        )
        del torchscript_module
        gc.collect()
    return coreml_model
 def get_out_path(submodule_name, model_name):
    from folder_paths import get_folder_paths
    fname = f"{model_name}_{submodule_name}.mlpackage"
    unet_path = get_folder_paths(submodule_name)[0]
    out_path = os.path.join(unet_path, fname)
    return out_path
 def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape, scheduler):
    sample_unet_inputs = dict(
        [
            ("sample", torch.rand(*sample_shape)),
            (
                "timestep",
                torch.tensor([scheduler.timesteps[0].item()] * batch_size).to(
                    torch.float32
                ),
            ),
            ("encoder_hidden_states", torch.rand(*encoder_hidden_states_shape)),
            ("timestep_cond", torch.randn(batch_size, 256).to(torch.float32)),
        ]
    )
    return sample_unet_inputs
 def get_unet_inputs_spec(sample_unet_inputs):
    sample_unet_inputs_spec = {
        k: (v.shape, v.dtype) for k, v in sample_unet_inputs.items()
    }
    return sample_unet_inputs_spec
 def add_cnet_support(sample_shape, reference_unet):
    additional_residuals_shapes = []
    batch_size = sample_shape[0]
    h, w = sample_shape[2:]
    # conv_in
    out_h, out_w = conv2d_output_shape(
        h,
        w,
        reference_unet.conv_in,
    )
    additional_residuals_shapes.append(
        (batch_size, reference_unet.conv_in.out_channels, out_h, out_w)
    )
    # down_blocks
    for down_block in reference_unet.down_blocks:
        additional_residuals_shapes += [
            (batch_size, resnet.out_channels, out_h, out_w)
            for resnet in down_block.resnets
        ]
        if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
            for downsampler in down_block.downsamplers:
                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
            additional_residuals_shapes.append(
                (
                    batch_size,
                    down_block.downsamplers[-1].conv.out_channels,
                    out_h,
                    out_w,
                )
            )
    # mid_block
    additional_residuals_shapes.append(
        (batch_size, reference_unet.mid_block.resnets[-1].out_channels, out_h, out_w)
    )
    additional_inputs = {}
    for i, shape in enumerate(additional_residuals_shapes):
        sample_residual_input = torch.rand(*shape)
        additional_inputs[f"additional_residual_{i}"] = sample_residual_input
    return additional_inputs
 def convert(
    out_path: str,
    batch_size: int = 1,
    sample_size: tuple[int, int] = (64, 64),
    controlnet_support: bool = False,
    lora_paths: list[str] = None,
 ):
    lora_paths = lora_paths or []
    coreml_unet, ref_unet = get_unets()
    for lora_path in lora_paths:
        lora_sd, network_alphas = LoraLoaderMixin.lora_state_dict(lora_path)
        LoraLoaderMixin.load_lora_into_unet(lora_sd, network_alphas, ref_unet)
        ref_unet.fuse_lora()
    sample_shape = (
        batch_size,  # B
        ref_unet.config.in_channels,  # C
        sample_size[0],  # H
        sample_size[1],  # W
    )
    encoder_hidden_states_shape = get_encoder_hidden_states_shape(
        ref_unet.config, batch_size
    )
    scheduler = get_scheduler()
    sample_inputs = get_sample_input(
        batch_size, encoder_hidden_states_shape, sample_shape, scheduler
    )
    if controlnet_support:
        sample_inputs |= add_cnet_support(sample_shape, ref_unet)
    sample_inputs_spec = get_unet_inputs_spec(sample_inputs)
    logger.info(f"Sample UNet inputs spec: {sample_inputs_spec}")
    logger.info("JIT tracing..")
    traced_unet = torch.jit.trace(
        coreml_unet, example_inputs=list(sample_inputs.values())
    )
    logger.info("Done.")
    coreml_sample_inputs = get_coreml_inputs(sample_inputs)
    coreml_unet = convert_to_coreml(
        "unet", traced_unet, coreml_sample_inputs, ["noise_pred"], out_path
    )
    del traced_unet
    gc.collect()
    coreml_unet.save(out_path)
    logger.info(f"Saved unet into {out_path}")
 if __name__ == "__main__":
    h = 512
    w = 512
    sample_size = (h // 8, w // 8)
    batch_size = 4
    cn_support_str = "_cn" if True else ""
    out_name = f"{MODEL_NAME}_{batch_size}x{w}x{h}{cn_support_str}"
    out_path = get_out_path("unet", f"{out_name}")
    if not os.path.exists(out_path):
        convert(out_path=out_path, sample_size=sample_size, batch_size=batch_size)
@@ -1,70 +0,0 @@
 import os
 from coremltools import ComputeUnit
 from coreml_suite import COREML_NODE
 from coreml_suite.coreml_model import CoreMLModel
 class COREML_CONVERT_LCM(COREML_NODE):
    """Converts a LCM model to Core ML."""
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "height": ("INT", {"default": 512, "min": 512, "max": 768, "step": 8}),
                "width": ("INT", {"default": 512, "min": 512, "max": 768, "step": 8}),
                "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                "compute_unit": (
                    [
                        ComputeUnit.CPU_AND_NE.name,
                        ComputeUnit.CPU_AND_GPU.name,
                        ComputeUnit.ALL.name,
                        ComputeUnit.CPU_ONLY.name,
                    ],
                ),
                "controlnet_support": ("BOOLEAN", {"default": False}),
            }
        }
    RETURN_TYPES = ("COREML_UNET",)
    RETURN_NAMES = ("coreml_model",)
    FUNCTION = "convert"
    def convert(self, height, width, batch_size, compute_unit, controlnet_support):
        """Converts a LCM model to Core ML.
        Args:
            height (int): Height of the target image.
            width (int): Width of the target image.
            batch_size (int): Batch size.
            compute_unit (str): Compute unit to use when loading the model.
        Returns:
            coreml_model: The converted Core ML model.
        The converted model is also saved to "models/unet" directory and
        can be loaded with the "LCMCoreMLLoaderUNet" node.
        """
        from coreml_suite.lcm import converter as lcm_converter
        h = height
        w = width
        sample_size = (h // 8, w // 8)
        batch_size = batch_size
        cn_support_str = "_cn" if controlnet_support else ""
        out_name = f"{lcm_converter.MODEL_NAME}_{batch_size}x{w}x{h}{cn_support_str}"
        out_path = lcm_converter.get_out_path("unet", f"{out_name}")
        if not os.path.exists(out_path):
            lcm_converter.convert(
                out_path=out_path,
                sample_size=sample_size,
                batch_size=batch_size,
                controlnet_support=controlnet_support,
            )
        return (CoreMLModel(out_path, compute_unit),)
@@ -1,98 +0,0 @@
 from diffusers import UNet2DConditionModel
 from diffusers.models.embeddings import TimestepEmbedding
 class UNet2DConditionModelLCM(UNet2DConditionModel):
    def __init__(
        self,
        time_cond_proj_dim=None,
        **kwargs,
    ):
        super().__init__(**kwargs)
        timestep_input_dim = self.config.block_out_channels[0]
        time_embed_dim = self.config.block_out_channels[0] * 4
        time_embedding = TimestepEmbedding(
            timestep_input_dim, time_embed_dim, cond_proj_dim=time_cond_proj_dim
        )
        self.time_embedding = time_embedding
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond,
        *additional_residuals,
    ):
        # 0. Project (or look-up) time embeddings
        t_emb = self.time_proj(timestep)
        emb = self.time_embedding(t_emb, timestep_cond)
        # 1. center input if necessary
        if self.config.center_input_sample:
            sample = 2 * sample - 1.0
        # 2. pre-process
        sample = self.conv_in(sample)
        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if (
                hasattr(downsample_block, "attentions")
                and downsample_block.attentions is not None
            ):
                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
            down_block_res_samples += res_samples
        if additional_residuals:
            new_down_block_res_samples = ()
            for i, down_block_res_sample in enumerate(down_block_res_samples):
                down_block_res_sample = down_block_res_sample + additional_residuals[i]
                new_down_block_res_samples += (down_block_res_sample,)
            down_block_res_samples = new_down_block_res_samples
        # 4. mid
        sample = self.mid_block(
            sample, emb, encoder_hidden_states=encoder_hidden_states
        )
        if additional_residuals:
            sample = sample + additional_residuals[-1]
        # 5. up
        for upsample_block in self.up_blocks:
            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[
                : -len(upsample_block.resnets)
            ]
            if (
                hasattr(upsample_block, "attentions")
                and upsample_block.attentions is not None
            ):
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples
                )
        # 6. post-process
        sample = self.conv_norm_out(sample)
        sample = self.conv_act(sample)
        sample = self.conv_out(sample)
        return (sample,)
@@ -1,8 +0,0 @@
 from enum import Enum
 class ModelVersion(Enum):
    SD15 = "sd15"
    SDXL = "sdxl"
    SDXL_REFINER = "sdxl_refiner"
    LCM = "lcm"
@@ -4,16 +4,9 @@ from coremltools import ComputeUnit
 import folder_paths
 from coreml_suite import COREML_NODE
 from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
 from coreml_suite.coreml_model import CoreMLModel
 from coreml_suite.core.naming import (
    QUANT_NBITS_VALUES,
    compose_out_name,
    lora_names_from_params,
 )
 from coreml_suite.lcm.utils import add_lcm_model_options, lcm_patch, is_lcm
 from coreml_suite.logger import logger
 from coreml_suite.model_version import ModelVersion
 from nodes import KSampler, LoraLoader, KSamplerAdvanced
 from coreml_suite.models import (
@@ -24,6 +17,26 @@ from coreml_suite.models import (
 )
 def _discover(fn_name, fallback):
    """Populate a converter dropdown from coreml_diffusion's discovery API.
    Fails soft: if the package is missing, too old to expose ``fn_name``, or
    errors, the node still registers with the fallback list instead of vanishing
    from the menu. Evaluated on every INPUT_TYPES call, so installing a newer
    coreml_diffusion surfaces new conversion types with no Suite change.
    """
    try:
        import coreml_diffusion
        return getattr(coreml_diffusion, fn_name)()
    except Exception as exc:  # missing/old package, import error, etc.
        logger.warning(
            f"coreml_diffusion.{fn_name} unavailable ({exc}); "
            f"using fallback {fallback}"
        )
        return fallback
 class CoreMLSampler(COREML_NODE, KSampler):
    @classmethod
    def INPUT_TYPES(s):
@@ -212,24 +225,26 @@ class CoreMLModelAdapter(COREML_NODE):
 class CoreMLConverter(COREML_NODE):
-    """Converts a LCM model to Core ML."""
+    """Converts a Stable Diffusion checkpoint (UNet) to Core ML.
    The model version (SD15 / SDXL / SDXL refiner / LCM) is auto-detected from
    the checkpoint's architecture, so there is no version dropdown — one node
    converts every supported family, including full-distill LCM.
    """
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "ckpt_name": (folder_paths.get_filename_list("checkpoints"),),
                "model_version": (
                    [
                        ModelVersion.SD15.name,
                        ModelVersion.SDXL.name,
                    ],
                ),
                "height": ("INT", {"default": 512, "min": 8, "step": 8}),
                "width": ("INT", {"default": 512, "min": 8, "step": 8}),
                "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                "attention_implementation": (
-                    list(ATTENTION_IMPLEMENTATIONS),
+                    _discover(
                        "list_attention_impls",
                        ["SPLIT_EINSUM", "SPLIT_EINSUM_V2", "ORIGINAL"],
                    ),
                ),
                "compute_unit": (
                    [
@@ -247,7 +262,10 @@ class CoreMLConverter(COREML_NODE):
                # omits any `required` input. When omitted it defaults to
                # "none", identical to unquantized behavior and filename, so
                # existing cached .mlpackages still resolve.
-                "quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),
+                "quantize_nbits": (
                    _discover("list_quant_modes", ["none", "8", "6", "4"]),
                    {"default": "none"},
                ),
                "lora_params": ("LORA_PARAMS",),
            },
        }
@@ -259,7 +277,6 @@ class CoreMLConverter(COREML_NODE):
    def convert(
        self,
        ckpt_name,
        model_version,
        height,
        width,
        batch_size,
@@ -269,9 +286,11 @@ class CoreMLConverter(COREML_NODE):
        quantize_nbits="none",
        lora_params=None,
    ):
-        """Converts a LCM model to Core ML.
+        """Converts a checkpoint's UNet to Core ML.
        Args:
            ckpt_name (str): Checkpoint to convert; its model version is
                auto-detected from the weights.
            height (int): Height of the target image.
            width (int): Width of the target image.
            batch_size (int): Batch size.
@@ -281,10 +300,8 @@ class CoreMLConverter(COREML_NODE):
            coreml_model: The converted Core ML model.
        The converted model is also saved to "models/unet" directory and
-        can be loaded with the "LCMCoreMLLoaderUNet" node.
+        can be loaded with the "Load Core ML UNet" node.
        """
        model_version = ModelVersion[model_version]
        lora_params = lora_params or {}
        lora_params = [(k, v[0]) for k, v in lora_params.items()]
        lora_params = sorted(lora_params, key=lambda lora: lora[0])
@@ -293,14 +310,16 @@ class CoreMLConverter(COREML_NODE):
        h = height
        w = width
        sample_size = (h // 8, w // 8)
-        out_name = compose_out_name(
+        import coreml_diffusion
        out_name = coreml_diffusion.compose_out_name(
            ckpt_name=ckpt_name,
            batch_size=batch_size,
            width=w,
            height=h,
            controlnet_support=controlnet_support,
            attention_implementation=attention_implementation,
-            lora_names=lora_names_from_params(lora_params),
+            lora_names=coreml_diffusion.lora_names_from_params(lora_params),
            quantize_nbits=quantize_nbits,
        )
@@ -311,13 +330,14 @@ class CoreMLConverter(COREML_NODE):
        logger.info(f"Attention implementation: {attention_implementation}")
        if lora_params:
-            logger.info(f"LoRAs used:")
+            logger.info("LoRAs used:")
            for lora_param in lora_params:
                logger.info(f"  {lora_param[0]} - strength: {lora_param[1]}")
-        from coreml_suite import converter
+        # Resolve the ComfyUI models/unet path here (a node concern); the package
-
+        # takes the output path as an injected argument.
-        unet_out_path = converter.get_out_path("unet", f"{out_name}")
+        unet_path = folder_paths.get_folder_paths("unet")[0]
        unet_out_path = os.path.join(unet_path, f"{out_name}_unet.mlpackage")
        ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
        config_filename = ckpt_name.split(".")[0] + ".yaml"
@@ -325,10 +345,10 @@ class CoreMLConverter(COREML_NODE):
        if config_path:
            logger.info(f"Using config file {config_path}")
-        converter.convert(
+        coreml_diffusion.convert(
-            ckpt_path=ckpt_path,
+            ckpt_path,
-            model_version=model_version,
+            None,  # model_version auto-detected from the checkpoint
-            unet_out_path=unet_out_path,
+            unet_out_path,
            sample_size=sample_size,
            batch_size=batch_size,
            controlnet_support=controlnet_support,
@@ -0,0 +1,93 @@
 # Conversion
 ## Conversion is the only supported path
 You always start from a Stable Diffusion checkpoint (`.safetensors` / `.ckpt`)
 and convert it with the **Convert Checkpoint to Core ML** node. The model
 version (SD1.5, SDXL, SDXL refiner, full-distill LCM) is auto-detected from the
 checkpoint. Pre-converted Core ML models from elsewhere are not supported,
 because:
 - The suite uses its own input **dimensions**, **naming convention**, and
  **metadata**, all produced by the
  [coreml-diffusion](https://github.com/aszc-dev/coreml-diffusion) package.
 - Apple's `ml-stable-diffusion` (which most community Core ML models target) is
  effectively obsolete, and the layouts differ (see the 2.0.0
  `encoder_hidden_states` change in the README).
 - Conversion is cheap and one-time, so there is no value in maintaining
  backwards compatibility with foreign formats.
 The output is always a **`.mlpackage`**. The suite no longer compiles to
 `.mlmodelc` (it didn't work with the inference backend), so there is **no Xcode
 or `coremlcompiler` dependency**.
 ## One-time conversion and name-based caching
 Conversion runs **once**, not on every queue. The converter encodes all
 conversion parameters into the output filename (via `coreml_diffusion.compose_out_name`,
 called in `coreml_suite/nodes.py`):
 - checkpoint name, `batch_size`, `width`, `height`
 - `controlnet_support`, `attention_implementation`
 - baked LoRA names, `quantize_nbits`
 The result is written as `<encoded-name>_unet.mlpackage` in `models/unet`. If a
 file with that name already exists, it is reused and conversion is skipped. Change
 any parameter → new name → new conversion; keep them the same → the cached model
 is loaded instantly.
 This is why the recommended workflow is **convert once, then load**: run the
 converter a single time, then in day-to-day use load the `.mlpackage` with the
 **Load Core ML UNet** node. (You can also leave the converter node in the graph;
 it short-circuits to the cached file.)
 > [!NOTE]
 > The converter relies on the filename to decide whether to reconvert. If you
 > rename the `.mlpackage`, it will be converted again. You can otherwise rename
 > it freely if the auto-generated name is too long.
 ## Quantization
 The converter node accepts an optional `quantize_nbits` dropdown that runs
 k-means weight palettization (`coremltools.optimize.coreml.palettize_weights`) on
 the UNet before saving.
 Values: `none` (default — no quantization, identical output and filename to
 before), `8`, `6`, `4`. The number is appended to the `.mlpackage` stem as
 `_q<bits>`, so quantized and unquantized variants coexist on disk and in cache.
 ### SD1.5 1×512×512 SPLIT_EINSUM tradeoffs (M2 Pro, ANE)
 Measured with 20 UNet forward passes at a fixed seed:
 | nbits | size (MB) | size vs none | fwd median (ms) | PSNR vs `none` (dB) |
 |---|---:|---:|---:|---:|
 | none | 1641 | 1.000 | 197.1 | — |
 | 8    |  822 | 0.501 | 186.6 | 53.5 |
 | 6    |  617 | 0.376 | 183.0 | 40.2 |
 | 4    |  412 | 0.251 | 179.8 | 27.5 |
 PSNR here is computed on the raw `noise_pred` output of a single UNet forward at a
 fixed seed, not on the final decoded image — it isolates quantization drift from
 sampler/VAE noise. Final-image PSNR is comfortably higher (the sampler averages
 over many steps).
 ### Recommended settings per chip / RAM
 - **8 GB (M1/M2 base):** `nbits=4`. ~4× smaller, still loads, 27 dB is visually
  identical at SD1.5 sizes.
 - **16 GB (M1/M2/M3 Pro):** `nbits=6` — the sweet spot, ~2.7× smaller, 40 dB, no
  perceptible quality drop.
 - **32 GB+ (Max / Ultra):** `nbits=8` for a safety margin, or `none` for
  bit-identical output (golden testing).
 The default stays `none`, so existing workflows produce byte-for-byte identical
 output.
 ## Where conversion lives
 The conversion engine was extracted into the standalone
 [coreml-diffusion](https://github.com/aszc-dev/coreml-diffusion) PyPI package.
 The nodes in this suite resolve ComfyUI paths and call into it; node names,
 inputs, and outputs are unchanged, so the split has effectively no user-facing
 impact beyond `pip install` pulling one more dependency.
@@ -0,0 +1,77 @@
 # FAQ
 ## What's the difference between ANE, GPU, and MPS, and which do I pick?
 ANE is the Neural Engine (Core ML only), GPU is the Metal GPU (Core ML or
 PyTorch), MPS is PyTorch's GPU backend. This suite uses **Core ML compute units
 only** and never touches MPS. Short answer: SD1.5 at 512×512 → convert
 `SPLIT_EINSUM`, load `CPU_AND_NE`; larger sizes or SDXL → convert `ORIGINAL`,
 load `CPU_AND_GPU`. Full reasoning: [hardware](hardware.md).
 ## Do I still need `PYTORCH_ENABLE_MPS_FALLBACK=1`?
 Not for these nodes — Core ML inference doesn't use PyTorch MPS. It may still
 matter for other parts of your ComfyUI graph, but it has no effect on Core ML
 sampling.
 ## Why is my Core ML SDXL workflow no faster than the default nodes?
 Because **SDXL can't run on the ANE** — the speedup comes from the Neural Engine,
 and SDXL falls back to the GPU, running at roughly MPS-equivalent speed. This is a
 known limitation, not a misconfiguration. The ANE benefit is real for SD1.5. See
 [limitations](limitations.md).
 ## Where do I get Core ML models?
 You convert them yourself — that's the only supported path. See
 [conversion](conversion.md). Downloaded Core ML models (e.g. coreml-community) use
 different dimensions/metadata and are not supported.
 ## Is conversion run every time I queue, or once?
 Once. Parameters are encoded in the output filename, so an already-converted model
 is reused and conversion is skipped. Convert once, then load the `.mlpackage`. See
 [conversion → caching](conversion.md#one-time-conversion-and-name-based-caching).
 ## Does a converted model produce the same output as the original?
 With the default `quantize_nbits = none`, the converted UNet output matches the
 source within numerical rounding (the golden test in `tests/m2/test_golden_image.py`
 gates on PSNR ≥ 20 dB on the decoded image). Quantization (`8`/`6`/`4`) introduces
 measured, bounded drift — see the [PSNR table](conversion.md#quantization). For
 bit-identical output, keep `none`.
 ## Are `.mlpackage` models safe to use?
 `.mlpackage` is a declarative Core ML model format — it carries weights and a
 compute graph, not arbitrary executable code or Python pickle, so its safety
 profile is comparable to `safetensors`. In practice this matters little here,
 since the only supported models are ones you convert locally from your own
 checkpoints.
 ## Are LoRAs reliable?
 Partially. Some LoRAs convert cleanly; others produce poor or broken output —
 there's no firm rule, so test per-LoRA. LoRA weights and `strength_model` are
 baked in at conversion and can't be changed afterward; for some LCM-LoRA cases the
 [Core ML Adapter](nodes.md#core-ml-adapter-experimental-coremlmodeladapter) path
 is more reliable. Treat LoRA support as experimental. See
 [troubleshooting](troubleshooting.md).
 ## Does the experimental Adapter cost performance vs the Core ML Sampler?
 Yes, a little. The Adapter wraps the model in a ComfyUI `ModelPatcher` so standard
 samplers work, which adds per-step interface overhead the native Core ML Sampler
 avoids. Use the native sampler unless you specifically need a `MODEL` (e.g.
 `ModelSamplingDiscrete` for LCM LoRAs).
 ## Which Python versions work?
 Python 3.12 or newer (`requires-python >=3.12`). Older 3.12 install failures
 came from the now-removed `ml-stable-diffusion` build, not from this suite.
 ## Long prompts crash my workflow
 Core ML has a hard **77-token** prompt limit and doesn't auto-chunk long prompts.
 Split the prompt across multiple CLIP Text Encode nodes and merge with Conditioning
 (Combine). See [troubleshooting](troubleshooting.md).
@@ -0,0 +1,95 @@
 # Hardware & Compute Units
 This page explains how the suite maps to Apple Silicon hardware, the difference
 between ANE, GPU, and MPS, and how to choose a compute unit and attention
 implementation.
 ## ANE vs GPU vs MPS
 Three terms get conflated:
 - **ANE (Apple Neural Engine)** — a dedicated ML accelerator on Apple Silicon.
  Only Core ML can target it; PyTorch cannot. This is the whole reason the suite
  exists.
 - **GPU** — the Metal GPU. Reachable both by Core ML (as a compute unit) and by
  PyTorch (via MPS).
 - **MPS (Metal Performance Shaders)** — PyTorch's GPU backend on macOS. This is
  the path standard ComfyUI nodes use.
 **This suite uses Core ML compute units only — it never runs the UNet through
 PyTorch/MPS.** Consequently `PYTORCH_ENABLE_MPS_FALLBACK` has no effect on these
 nodes. It may still matter for the rest of your ComfyUI graph (CLIP, VAE,
 samplers on non-Core ML models), but not for Core ML inference itself.
 Rough performance picture (SD1.5, maintainer- and user-reported):
 - ANE is meaningfully faster than MPS — on the order of **50–100%** for SD1.5.
 - Core ML on the GPU is only marginally faster than PyTorch/MPS.
 So the speedup comes from the Neural Engine, which means it depends on being able
 to actually run on the ANE (see [attention implementations](#attention-implementations)
 and the [SDXL caveat](#sdxl-and-the-ane)).
 ## Compute units
 The **compute unit** is set on the loader/converter node and tells Core ML which
 hardware to use. It is applied when the model is loaded
 (`coreml_suite/coreml_model.py:22`), not during conversion.
 | Value | Hardware | Best paired with |
 |---|---|---|
 | `CPU_AND_NE` (default) | CPU + Neural Engine | `SPLIT_EINSUM` / `SPLIT_EINSUM_V2` |
 | `CPU_AND_GPU` | CPU + Metal GPU | `ORIGINAL` |
 | `CPU_ONLY` | CPU only | fallback / debugging |
 | `ALL` | all available hardware | rarely optimal — see below |
 Notes:
 - Every option includes the CPU; there is no GPU-and-ANE-without-CPU combination.
 - `NE` in `CPU_AND_NE` is the Neural Engine (Apple's enum spells it `NE`, not
  `ANE`).
 - **`CPU_AND_NE` is often faster than `ALL`.** Letting Core ML use everything can
  be *slower* on non-Max chips, where memory bandwidth is the bottleneck. Try
  `CPU_AND_NE` first for SD1.5.
 ## Attention implementations
 Chosen at conversion time on the **Convert Checkpoint to Core ML** node. It
 decides whether the model can run on the ANE:
 - **`SPLIT_EINSUM`** — ANE-friendly attention. Use for the Neural Engine.
 - **`SPLIT_EINSUM_V2`** — a variant; in practice ≈ `SPLIT_EINSUM` for most users.
 - **`ORIGINAL`** — standard attention. Runs on the GPU, not the ANE.
 The implementation and the compute unit must agree: a `SPLIT_EINSUM` model wants
 `CPU_AND_NE`; an `ORIGINAL` model wants `CPU_AND_GPU`.
 ## Which should I pick?
 | Scenario | Attention | Compute unit |
 |---|---|---|
 | SD1.5 at 512×512 | `SPLIT_EINSUM` | `CPU_AND_NE` |
 | SD1.5 at larger sizes (e.g. 768) | `ORIGINAL` | `CPU_AND_GPU` |
 | SDXL / SDXL Turbo | `ORIGINAL` | `CPU_AND_GPU` |
 ### Resolution crossover
 ANE shines at small latents; the GPU scales better as resolution grows. In user
 benchmarks:
 - At **512×512**, ANE + `SPLIT_EINSUM` wins by roughly **10%** over the GPU path.
 - At **768×512**, GPU + `ORIGINAL` pulls ahead by roughly **10%**, and the larger
  image is about 2× slower overall.
 If you mostly work at 512×512, convert with `SPLIT_EINSUM` and load on
 `CPU_AND_NE`. If you routinely go larger, an `ORIGINAL` + GPU model may be
 faster.
 ### SDXL and the ANE
 SDXL (and SDXL Turbo) **cannot run on the ANE** — the dual-text-encoder UNet
 exceeds what the Neural Engine path supports. SDXL therefore runs at roughly
 MPS-equivalent speed with no ANE speedup. If a Core ML SDXL workflow feels no
 faster than the standard nodes, this is why. Convert SDXL with `ORIGINAL` and
 load with `CPU_AND_GPU` or `CPU_ONLY`. See
 [limitations](limitations.md) for the full picture.
@@ -0,0 +1,53 @@
 # Limitations & Support Matrix
 ## Support matrix
 | Feature | Status | Notes |
 |---|---|---|
 | SD1.5 | ✅ Full | ANE via `SPLIT_EINSUM`; the primary, fastest path |
 | SDXL / SDXL Turbo | ⚠️ Partial | GPU only (no ANE), no speedup; possible quality loss vs source. Don't run Turbo at 1024² |
 | SD2.1 | ❌ Unsupported | |
 | Inpainting checkpoints (9-channel) | ❌ Unsupported | |
 | ControlNet | ✅ Supported | Convert the checkpoint with `controlnet_support = True` |
 | LoRA | ⚠️ Experimental | Inconsistent per-LoRA; baked at conversion, immutable afterward |
 | LCM | ⚠️ Experimental | Full-distill LCM checkpoints auto-detected by the converter |
 | SVD | ❌ Not supported | |
 | AnimateDiff | ❌ Not supported | Motion modules need pre-conversion injection; not feasible today |
 | IPAdapter | ❌ Not supported | Needs a real `MODEL` the Core ML wrapper can't provide |
 | Core ML Adapter | ⚠️ Experimental | Works for many nodes; fails for merges/IPAdapter/etc. |
 ## Fixed input/output shapes
 A Core ML model is converted for one specific resolution and batch size. To work
 at a different size, re-convert with the new width/height (conversion is cheap and
 cached by name). This is also why detailers and latent-upscale workflows that
 rescale mid-graph break — see [troubleshooting](troubleshooting.md).
 There is experimental support for flexible shapes via
 [EnumeratedShapes](https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html#select-from-predetermined-shapes),
 but it is **much slower** — user benchmarks show roughly **5×** the per-iteration
 time on every run, not just the first. Fixed-shape models per resolution are the
 practical choice.
 ## SDXL on the Neural Engine
 SDXL and SDXL Turbo cannot run on the ANE — the dual-text-encoder UNet exceeds the
 supported Neural Engine path. They run on the GPU at roughly MPS-equivalent speed,
 so Core ML offers no speed advantage for SDXL, and converted output may look
 degraded versus the safetensors original (an upstream conversion artifact). Use
 `ORIGINAL` + `CPU_AND_GPU`. See [hardware](hardware.md).
 ## Experimental Core ML Adapter
 The Adapter wraps a Core ML model to look like a standard ComfyUI `MODEL`, which
 covers many standard and custom nodes. But it can't fully emulate a real model:
 operations that need genuine `MODEL` internals — model merges, IPAdapter, some
 LoRA flows, detailers — generally won't work, and the model's
 fixed input shapes aren't validated, so mismatches error at runtime. Prefer the
 native Core ML Sampler when you don't need the `MODEL` type.
 ## Prompt length
 Core ML enforces a hard 77-token prompt limit with no auto-chunking. Split long
 prompts across multiple CLIP Text Encode nodes and merge with Conditioning
 (Combine).
@@ -0,0 +1,157 @@
 # Node Reference
 All nodes live in the **Core ML Suite** category. Right-click the canvas →
 **Add Node → Core ML Suite**, or double-click and search.
 | Display name | Class | Purpose |
 |---|---|---|
 | Load Core ML UNet | `CoreMLUNetLoader` | Load a converted `.mlpackage` |
 | Core ML Sampler | `CoreMLSampler` | Sample (KSampler-style) |
 | Core ML Sampler (Advanced) | `CoreMLSamplerAdvanced` | Sample (KSamplerAdvanced-style) |
 | Core ML Adapter (Experimental) | `CoreMLModelAdapter` | Wrap as a standard `MODEL` |
 | Load LoRA to use with Core ML | `Core ML LoRA Loader` | Bake LoRA(s) at conversion |
 | Convert Checkpoint to Core ML | `Core ML Converter` | Convert a checkpoint |
 ---
 ## Load Core ML UNet (`CoreMLUNetLoader`)
 ![Load Core ML UNet](../assets/unet_loader.png?raw=true)
 Loads a converted `.mlpackage` from `models/unet` and outputs a `coreml_model`
 for the samplers. Only `.mlpackage` files are listed — this suite no longer uses
 `.mlmodelc`.
 - **Inputs**
  - `coreml_name` — the `.mlpackage` to load from `models/unet`.
  - `compute_unit` — hardware to run on: `CPU_AND_NE` (default), `CPU_AND_GPU`,
    `CPU_ONLY`, `ALL`. See [hardware](hardware.md).
 - **Output**
  - `coreml_model` — for the Core ML Sampler or Adapter.
 ---
 ## Core ML Sampler (`CoreMLSampler`)
 ![Core ML Sampler](../assets/sampler.png?raw=true)
 Generates a latent from a Core ML model. Behaves like the standard KSampler and
 outputs a `LATENT` you can decode or feed downstream.
 - **Inputs**
  - `coreml_model` — output of the loader or a converter.
  - `latent_image` *(optional)* — must match the model's input size. If omitted,
    a suitable empty latent is created. Provide one for img2img.
  - `negative` *(optional)* — required for normal models; optional for LCM.
  - Remaining inputs (`seed`, `steps`, `cfg`, `sampler_name`, `scheduler`,
    `positive`, `denoise`) match the KSampler.
 - **Output**
  - `LATENT` — decode with a VAE Decode, or use downstream.
 ---
 ## Core ML Sampler (Advanced) (`CoreMLSamplerAdvanced`)
 The KSamplerAdvanced counterpart of the Core ML Sampler — same Core ML input,
 plus the advanced sampling controls. Use it for partial denoising, fixed noise,
 and multi-stage (e.g. SDXL base → refiner) workflows.
 - **Inputs**
  - `coreml_model` — output of the loader or a converter.
  - `add_noise`, `noise_seed`, `start_at_step`, `end_at_step`,
    `return_with_leftover_noise` — as in KSamplerAdvanced.
  - `steps`, `cfg`, `sampler_name`, `scheduler`, `positive` — as usual.
  - `latent_image` *(optional)*, `negative` *(optional, required for non-LCM)*.
 - **Output**
  - `LATENT`.
 ---
 ## Core ML Adapter (Experimental) (`CoreMLModelAdapter`)
 ![Core ML Adapter](../assets/adapter.png?raw=true)
 Wraps a Core ML model so it presents as a standard ComfyUI `MODEL`, letting you
 feed it to the normal KSampler and many other nodes (e.g. `ModelSamplingDiscrete`
 for LCM LoRAs).
 - **Input**
  - `coreml_model`.
 - **Output**
  - `MODEL` — a Core ML model wrapped as a ComfyUI model.
 > [!NOTE]
 > Experimental. The wrapper presents a `MODEL` interface but cannot fully
 > emulate one — model merges, IPAdapter, and similar advanced uses generally
 > won't work, and the model's fixed input shapes are not validated, so mismatched
 > inputs error at runtime. The native Core ML Sampler is faster when you don't
 > need the `MODEL` type. See the [FAQ](faq.md) and [limitations](limitations.md).
 ---
 ## Load LoRA to use with Core ML (`Core ML LoRA Loader`)
 ![LoRA Loader](../assets/lora_loader.png?raw=true)
 Collects LoRA name + `strength_model` to bake into the model at conversion, and
 applies the LoRA to CLIP (which is not part of the Core ML path). Chain multiple
 loaders for multiple LoRAs.
 Because a converted model is immutable, the baked weights and `strength_model`
 **cannot** be changed afterward — changing them means re-converting. `strength_clip`
 only affects CLIP and can be changed freely. After conversion, when loading with
 `CoreMLUNetLoader`, apply the same LoRAs to CLIP manually (see
 [workflows](workflows.md)).
 - **Inputs**
  - `lora_name`, `strength_model`, `strength_clip`.
  - `clip` — from `CLIPLoader` / `CheckpointLoaderSimple` or another LoRA loader.
  - `lora_params` *(optional)* — chain from another LoRA loader.
 - **Outputs**
  - `CLIP` — with the LoRA applied.
  - `lora_params` — pass to the converter or the next LoRA loader.
 > [!NOTE]
 > LoRA support is experimental and inconsistent — some LoRAs convert cleanly,
 > others produce poor results. Test per-LoRA. See [troubleshooting](troubleshooting.md).
 ---
 ## Convert Checkpoint to Core ML (`Core ML Converter`)
 ![Checkpoint Converter](../assets/checkpoint_converter.png?raw=true)
 Converts a checkpoint from `models/checkpoints` to a Core ML `.mlpackage` in
 `models/unet`. The model version (SD1.5, SDXL, SDXL refiner, or full-distill
 LCM) is auto-detected from the checkpoint's architecture — there is no version
 dropdown. The conversion parameters are encoded in the output name, so an
 already-converted model is reused instead of re-converted. See
 [conversion](conversion.md) for details.
 - **Inputs**
  - `ckpt_name` — checkpoint in `models/checkpoints`.
  - `height`, `width` — target image size; any positive multiple of 8 (default
    512). The model's input size is fixed at these values.
  - `batch_size` — default 1; raise to convert a batch-capable model.
  - `attention_implementation` — `SPLIT_EINSUM` / `SPLIT_EINSUM_V2` (ANE) or
    `ORIGINAL` (GPU). See [hardware](hardware.md).
  - `compute_unit` — used only when loading the result; does not affect
    conversion.
  - `controlnet_support` — set `True` to make the model usable with ControlNet
    (default `False`).
  - `quantize_nbits` *(optional)* — `none` (default), `8`, `6`, `4`. See
    [conversion → quantization](conversion.md#quantization).
  - `lora_params` *(optional)* — from the LoRA loader, to bake LoRAs in.
 - **Output**
  - `coreml_model`.
 > [!NOTE]
 > Some checkpoints need a custom config `.yaml`. Place it in `models/configs`
 > named like the checkpoint (e.g. `juggernaut.safetensors` →
 > `juggernaut.yaml`); it is loaded automatically during conversion.
 > [!NOTE]
 > Full-distill LCM checkpoints (e.g.
 > [LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7)) are
 > detected and converted like any other checkpoint. When sampling an LCM model,
 > set `sampler_name` to `lcm` and `scheduler` to `sgm_uniform`.
@@ -0,0 +1,86 @@
 # Troubleshooting
 ## `Expected shape … got …` / latent size mismatch
 The most common error. A Core ML model has **fixed** input dimensions — a model
 converted for 512×512 expects a 64×64 latent and rejects any other size (batch
 size is handled and doesn't matter; only width/height are fixed).
 **Fix:** set your Empty Latent (or upstream latent) to exactly the resolution the
 model was converted for, or re-convert at the size you want.
 ## Old `.mlmodelc` model, or `metadata.json` not found
 This suite no longer produces or loads `.mlmodelc`; the loader lists `.mlpackage`
 only. Models from an older version (or downloaded community models) with a
 `.mlmodelc` structure won't load.
 **Fix:** re-convert the checkpoint with **Convert Checkpoint to Core ML**. No
 Xcode or `coremlcompiler` is required — that dependency was removed.
 ## Prompt too long (`Expected size 154 but got 77`, or a crash)
 Core ML enforces a hard **77-token** prompt limit and does not auto-chunk like
 A1111/ComfyUI.
 **Fix:** split the prompt across multiple CLIP Text Encode nodes and merge them
 with **Conditioning (Combine)**.
 ## `cannot import name 'ModelSamplingDiscreteLCM'`
 A ComfyUI refactor renamed this symbol.
 **Fix:** update the suite (fixed in PR #29) and re-run
 `pip install -r requirements.txt`.
 ## LoRA loader `ImportError`
 `peft` became a required dependency.
 **Fix:** `pip install -r requirements.txt`. This recurs after ComfyUI-Manager
 updates if requirements aren't reinstalled.
 ## ControlNet has no effect
 ControlNet support is baked at conversion. If the checkpoint was converted with
 `controlnet_support = False`, ControlNet does nothing.
 **Fix:** re-convert with `controlnet_support = True`. The ControlNet model itself
 needs no conversion, and `.fp16.safetensors` vs `.safetensors` makes no
 difference.
 ## LoRAs produce garbage
 LoRA support is inconsistent — some work, some don't, with no firm rule. Test
 per-LoRA. For some LCM-LoRA setups, routing through the
 [Core ML Adapter](nodes.md#core-ml-adapter-experimental-coremlmodeladapter) is
 more reliable than the basic loader path. Remember weights are baked at conversion
 and can't be changed afterward.
 ## FaceDetailer / detailers error on size
 Detailers rescale latents internally (e.g. 512 → 1024), which breaks the model's
 fixed input shape. There is no workaround node — a Core ML model only accepts
 the resolution it was converted for.
 **Fix:** convert a second model at the detailer's internal resolution and use it
 for the detailing pass, or run the detailer with a standard (non–Core ML) model.
 ## Inpainting checkpoint errors (`tensor size 9 vs 4`)
 SD1.5 inpainting checkpoints use a 9-channel input and are **not supported**. This
 error is expected, not a bug.
 ## Errors mentioning `python_coreml_stable_diffusion` or `ml-stable-diffusion`
 You're on a stale install. That dependency was removed; old install scripts tried
 `pip install git+…/ml-stable-diffusion.git`, which fails on modern Python.
 **Fix:** reinstall the current suite (`pip install -r requirements.txt`, which
 pulls `coreml-diffusion` from PyPI).
 ## `all input tensors must be on the same device (mps:0 and cpu)` / ControlNet residual shape `(2,…) vs (1,…)`
 Old bugs that have been fixed.
 **Fix:** update to the latest version.
@@ -0,0 +1,103 @@
 # Example Workflows
 > [!NOTE]
 > The models referenced are examples — substitute your own. Every workflow
 > starts from a checkpoint you convert yourself (see [conversion](conversion.md));
 > there is no Core ML model to download.
 ## Basic txt2img
 Convert a SD1.5 checkpoint, then sample from it. CLIP and VAE come from standard
 ComfyUI nodes — either loaded separately or pulled from the checkpoint.
 1. Place a SD1.5 checkpoint in `models/checkpoints` (e.g.
   [v1-5-pruned-emaonly](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors)).
 2. **Convert Checkpoint to Core ML** → queue once → a `.mlpackage` lands in
   `models/unet`.
 3. **Load Core ML UNet** (or wire the converter output straight in) →
   **Core ML Sampler** → **VAE Decode**.
 **CLIP and VAE from the checkpoint:**
 ![Core ML UNet + checkpoint](../assets/unet+sampler+checkpoint.png?raw=true)
 **CLIP and VAE loaded separately** — use any SD1.5-compatible
 [CLIP](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/text_encoder/model.safetensors)
 and [VAE](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/vae/diffusion_pytorch_model.safetensors),
 placed in `models/clip` and `models/vae`:
 ![Core ML UNet + CLIP + VAE](../assets/unet+sampler+clip+vae.png?raw=true)
 ## ControlNet
 Convert the checkpoint with `controlnet_support = True`, then wire a standard
 ComfyUI ControlNet. The ControlNet model itself needs no conversion. Place it in
 `models/controlnet` (e.g.
 [control_v11p_sd15_scribble](https://huggingface.co/lllyasviel/control_v11p_sd15_scribble/blob/main/diffusion_pytorch_model.fp16.safetensors)).
 ![Core ML UNet + ControlNet](../assets/unet+sampler+controlnet.png?raw=true)
 ## Checkpoint conversion
 The minimal conversion graph. See
 [Convert Checkpoint to Core ML](nodes.md#convert-checkpoint-to-core-ml-core-ml-converter).
 ![Checkpoint converter](../assets/basic_conversion.png?raw=true)
 ## Conversion with LoRA
 Bake LoRA(s) into the model at conversion. Read the
 [LoRA caveats](nodes.md#load-lora-to-use-with-core-ml-core-ml-lora-loader) first
 — baked weights are immutable, and support is inconsistent per-LoRA.
 ![Checkpoint converter + LoRA](../assets/conversion+lora.png?raw=true)
 ## LCM LoRA conversion
 Chain multiple LoRA loaders to use several LoRAs with one model.
 > [!IMPORTANT]
 > Here the model goes through the **Core ML Adapter** and `ModelSamplingDiscrete`
 > into the standard ComfyUI KSampler (not the Core ML Sampler).
 > `ModelSamplingDiscrete` is required to sample LCM LoRAs correctly.
 ![Multiple LoRAs](../assets/conversion+lcm_lora.png?raw=true)
 ## Loading a model with baked LoRAs
 Load a model that already has LoRAs baked in. CLIP must be loaded separately and
 passed through the same LoRA nodes used at conversion. Since `lora_name` and
 `strength_model` are baked in, they need not be passed to the loader.
 > [!IMPORTANT]
 > As above, the model goes through the Core ML Adapter + `ModelSamplingDiscrete`
 > into the standard KSampler.
 ![Loader + LoRA](../assets/loader+lcm_lora.png?raw=true)
 ## LCM conversion with ControlNet
 Convert a full-distill LCM checkpoint (e.g.
 [LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7)) with
 the standard **Convert Checkpoint to Core ML** node — the LCM architecture is
 auto-detected. Use it with or without ControlNet. When sampling, set
 `sampler_name` to `lcm` and `scheduler` to `sgm_uniform`.
 ![LCM + ControlNet](../assets/lcm+controlnet.png?raw=true)
 ## SDXL Base + Refiner
 A basic SDXL graph. Add LoRAs and ControlNets as in the SD1.5 examples; the
 refiner step is optional.
 Models:
 [base + text encoders](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0),
 [refiner](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0),
 [VAE](https://huggingface.co/stabilityai/sdxl-vae).
 > [!IMPORTANT]
 > SDXL does not run on the ANE. Convert with `ORIGINAL` and load with
 > `CPU_AND_GPU` (or `CPU_ONLY`). If loading hangs on `CPU_AND_NE`, that is the
 > cause. See [limitations](limitations.md).
 ![SDXL](../assets/sdxl_conversion.png?raw=true)
@@ -1,27 +1,34 @@
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
 [project]
 name = "comfyui-coremlsuite"
 description = "This extension contains a set of custom nodes for ComfyUI that allow you to use Core ML models in your ComfyUI workflows."
-version = "2.0.2"
+version = "2.1.2"
 license = "MIT"
-requires-python = ">=3.12,<3.13"
+requires-python = ">=3.12"
 packages = [{ include = "coreml_suite" }]
 dependencies = [
-    "torch>=2.7,<2.8",
+    # torch is provided by the host (ComfyUI) and intentionally left unpinned
    # here: a hard torch cap would downgrade the host's torch and break its
    # torchvision/torchaudio ABI. coreml-diffusion pulls torch>=2.7 transitively.
    # >=0.1.6: model-version auto-detection (convert(model_version=None)) and the
    # dropped <3.13 Python cap (kept in sync with this package's requires-python).
    "coreml-diffusion>=0.1.6,<0.2",
    "coremltools>=9,<10",
    "numpy>=2,<3",
    "diffusers>=0.30",
    "peft>=0.13",
    "omegaconf>=2.3",
    "transformers>=4.44",
 ]
 [project.urls]
 Repository = "https://github.com/aszc-dev/ComfyUI-CoreMLSuite"
 [tool.hatch.build.targets.wheel]
 packages = ["coreml_suite"]
 [tool.comfy]
 PublisherId = "aszc-dev"
 DisplayName = "ComfyUI-CoreMLSuite"
-Icon = ""
+Icon = "https://raw.githubusercontent.com/aszc-dev/ComfyUI-CoreMLSuite/main/assets/snake.png"
 requires-comfyui = ">=0.3.27"
 [dependency-groups]
@@ -1,7 +1,4 @@
-torch>=2.7,<2.8
+coreml-diffusion>=0.1.4,<0.2
 coremltools>=9,<10
 numpy>=2,<3
 diffusers>=0.30
 peft>=0.13
 omegaconf>=2.3
 transformers>=4.44
@@ -107,7 +107,6 @@
  "10": {
    "inputs": {
      "ckpt_name": "dreamshaper_8.safetensors",
      "model_version": "SD15",
      "height": 512,
      "width": 512,
      "batch_size": 1,
@@ -1,41 +0,0 @@
 import platform
 import pytest
 import torch
 from diffusers.models.attention_processor import Attention, AttnProcessor
 from coreml_suite.conversion.attention import (
    SplitEinsumAttnProcessor,
    SplitEinsumV2AttnProcessor,
 )
 pytestmark = pytest.mark.skipif(
    platform.system() != "Darwin" or platform.machine() != "arm64",
    reason="Tier 1 requires macOS on Apple Silicon",
 )
@pytest.mark.parametrize(
    "processor",
    [
        SplitEinsumAttnProcessor(),
        SplitEinsumV2AttnProcessor(),
    ],
 )
 def test_split_einsum_processor_matches_diffusers_attention(processor):
    torch.manual_seed(0)
    reference = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
    reference.set_processor(AttnProcessor())
    candidate = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
    candidate.load_state_dict(reference.state_dict())
    candidate.set_processor(processor)
    hidden_states = torch.randn(2, 17, 32)
    encoder_hidden_states = torch.randn(2, 11, 32)
    expected = reference(hidden_states, encoder_hidden_states=encoder_hidden_states)
    actual = candidate(hidden_states, encoder_hidden_states=encoder_hidden_states)
    assert torch.allclose(actual, expected, atol=1e-5)
@@ -1,138 +0,0 @@
 """Tier 1 smoke: convert a synthetic micro-UNet through coremltools and load
 it back with CoreMLSuite's runtime CoreMLModel wrapper.
 Purpose: catch API breakage in coremltools *without* needing a real SD
 checkpoint, the ANE, or a converted .mlmodelc on disk.
 Runs in minutes on a hosted macOS-ARM runner (no Apple internal stuff).
 What it asserts:
  - coremltools.convert still accepts the call shape we use today
  - the resulting .mlpackage round-trips through CoreMLSuite's CoreMLModel
  - expected_inputs exposes the input names/shapes we declared
  - calling the model returns the named output (`noise_pred`)
 Auto-skips on non-Apple-Silicon hosts so Tier 0 CI on Linux ignores it.
 """
 import platform
 import shutil
 from types import SimpleNamespace
 import numpy as np
 import pytest
 import torch
 import torch.nn as nn
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 pytestmark = pytest.mark.skipif(
    platform.system() != "Darwin" or platform.machine() != "arm64",
    reason="Tier 1 requires macOS on Apple Silicon",
 )
 # Tiny shapes — large enough to exercise conv2d + linear + addition kernels in
 # coremltools, small enough that conversion finishes in seconds on CPU.
 SAMPLE_SHAPE = (1, 4, 8, 8)
 TIMESTEP_SHAPE = (1,)
 ENCODER_SHAPE = (1, 4, 64)  # native diffusers encoder_hidden_states (batch, tokens, hidden)
 OUT_NAME = "noise_pred"
 class TinyUNet(nn.Module):
    """Minimal UNet-shaped graph: conv -> add(time+context) -> conv.
    Not a real diffusion model. Just enough op variety to exercise the
    PyTorch -> MIL frontend in coremltools and confirm we can still wire
    the inputs/outputs the way CoreMLSuite's runtime expects.
    """
    def __init__(self):
        super().__init__()
        self.conv_in = nn.Conv2d(4, 8, kernel_size=3, padding=1)
        self.conv_out = nn.Conv2d(8, 4, kernel_size=3, padding=1)
        self.time_proj = nn.Linear(1, 8)
        self.text_proj = nn.Linear(64, 8)
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond=None,
        added_cond_kwargs=None,
        down_block_additional_residuals=None,
        mid_block_additional_residual=None,
        return_dict=True,
    ):
        h = self.conv_in(sample)
        t_emb = self.time_proj(timestep.unsqueeze(-1)).view(1, 8, 1, 1)
        c_emb = self.text_proj(encoder_hidden_states.mean(1)).view(1, 8, 1, 1)
        h = h + t_emb + c_emb
        return (self.conv_out(h),)
@pytest.fixture(scope="module")
 def tiny_mlpackage(tmp_path_factory):
    """Convert TinyUNet once per test session and reuse the .mlpackage."""
    import coremltools as ct
    torch.manual_seed(0)
    model = CoreMLUNetWrapper(
        TinyUNet().eval(),
        SimpleNamespace(name="SD15"),
    )
    example = (
        torch.randn(*SAMPLE_SHAPE),
        torch.randn(*TIMESTEP_SHAPE),
        torch.randn(*ENCODER_SHAPE),
    )
    traced = torch.jit.trace(model, example)
    mlmodel = ct.convert(
        traced,
        inputs=[
            ct.TensorType(name="sample", shape=SAMPLE_SHAPE, dtype=np.float16),
            ct.TensorType(name="timestep", shape=TIMESTEP_SHAPE, dtype=np.float16),
            ct.TensorType(name="encoder_hidden_states", shape=ENCODER_SHAPE, dtype=np.float16),
        ],
        outputs=[ct.TensorType(name=OUT_NAME, dtype=np.float16)],
        compute_units=ct.ComputeUnit.CPU_ONLY,
        compute_precision=ct.precision.FLOAT16,
        convert_to="mlprogram",
        minimum_deployment_target=ct.target.macOS13,
    )
    out_dir = tmp_path_factory.mktemp("tiny_unet")
    pkg_path = out_dir / "tiny.mlpackage"
    mlmodel.save(str(pkg_path))
    yield pkg_path
    shutil.rmtree(out_dir, ignore_errors=True)
 def test_coremltools_convert_round_trips_via_coreml_model(tiny_mlpackage):
    from coreml_suite.coreml_model import CoreMLModel
    model = CoreMLModel(str(tiny_mlpackage), "CPU_ONLY")
    # expected_inputs is the contract our wrappers depend on. Lock the shape
    # of the dict + a sample entry.
    expected = dict(model.expected_inputs)
    assert set(expected.keys()) == {"sample", "timestep", "encoder_hidden_states"}
    assert tuple(expected["sample"]["shape"]) == SAMPLE_SHAPE
    assert tuple(expected["timestep"]["shape"]) == TIMESTEP_SHAPE
    assert tuple(expected["encoder_hidden_states"]["shape"]) == ENCODER_SHAPE
    # Forward pass: drive the model the way CoreMLModelWrapper does.
    rng = np.random.default_rng(0)
    inputs = {
        "sample": rng.standard_normal(SAMPLE_SHAPE).astype(np.float16),
        "timestep": rng.standard_normal(TIMESTEP_SHAPE).astype(np.float16),
        "encoder_hidden_states": rng.standard_normal(ENCODER_SHAPE).astype(np.float16),
    }
    out = model(**inputs)
    assert isinstance(out, dict), f"unexpected output type: {type(out)}"
    assert OUT_NAME in out, f"missing output {OUT_NAME!r}; got {sorted(out)}"
    assert out[OUT_NAME].shape == SAMPLE_SHAPE, (
        f"output shape drift: got {out[OUT_NAME].shape}, expected {SAMPLE_SHAPE}"
    )
@@ -1,197 +0,0 @@
 """Characterization tests for the .mlpackage filename composition.
 The filename composition is the pure
 coreml_suite.core.naming.compose_out_name function. CoreMLConverter.convert
 calls it; testing the pure function avoids monkey-patching heavy converter
 internals just to capture the string.
 """
 import pytest
 from coreml_suite.core.naming import compose_out_name, lora_names_from_params
 # ---------- attention suffixes ----------------------------------------------
@pytest.mark.parametrize(
    "attn_name,suffix",
    [
        ("SPLIT_EINSUM", "se"),
        ("SPLIT_EINSUM_V2", "se2"),
        ("ORIGINAL", "orig"),
    ],
 )
 def test_attention_suffix(attn_name, suffix):
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation=attn_name,
    )
    assert out == f"dreamshaper_8_1x512x512_{suffix}"
 # ---------- batch / size ----------------------------------------------------
 def test_includes_batch_and_size():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=4, width=768, height=1024,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dreamshaper_8_4x768x1024_se"
 # ---------- ControlNet ------------------------------------------------------
 def test_appends_cn_suffix_when_controlnet_support_true():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dreamshaper_8_1x512x512_cn_se"
 # ---------- ckpt name massage -----------------------------------------------
 def test_drops_extension_at_first_period():
    out = compose_out_name(
        ckpt_name="my.checkpoint.v2.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "my_1x512x512_se"
 def test_replaces_spaces_with_underscores():
    out = compose_out_name(
        ckpt_name="dream shaper 8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
    )
    assert out == "dream_shaper_8_1x512x512_se"
 # ---------- LoRA suffixes ---------------------------------------------------
 def test_single_lora():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["epi_noiseoffset.safetensors"],
    )
    assert out == "dreamshaper_8_epi_noiseoffset_1x512x512_se"
 def test_multiple_loras_sorted():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["zoom.safetensors", "alpha.safetensors", "moody.safetensors"],
    )
    assert out == "dreamshaper_8_alpha_moody_zoom_1x512x512_se"
 def test_lora_plus_controlnet():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["a.safetensors"],
    )
    assert out == "dreamshaper_8_a_1x512x512_cn_se"
 # ---------- sdxl combinations -----------------------------------------------
 def test_sdxl_1024_original_gpu():
    out = compose_out_name(
        ckpt_name="sd_xl_base_1.0.safetensors",
        batch_size=1, width=1024, height=1024,
        controlnet_support=False,
        attention_implementation="ORIGINAL",
    )
    assert out == "sd_xl_base_1_1x1024x1024_orig"
 # ---------- lora_names_from_params helper ----------------------------------
 def test_lora_names_from_params_sorts_by_name():
    names = lora_names_from_params([
        ("zebra.safetensors", 1.0),
        ("apple.safetensors", 0.5),
        ("mango.safetensors", 0.7),
    ])
    assert names == ["apple.safetensors", "mango.safetensors", "zebra.safetensors"]
 def test_lora_names_from_params_empty_list():
    assert lora_names_from_params([]) == []
 # ---------- quantize_nbits suffix ------------------------------------------
 def test_quantize_nbits_none_appends_nothing():
    """'none' is the default and must keep the unquantized filename so
    existing cached .mlpackages still resolve."""
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        quantize_nbits="none",
    )
    assert out == "dreamshaper_8_1x512x512_se"
@pytest.mark.parametrize("nbits,suffix", [("4", "_q4"), ("6", "_q6"), ("8", "_q8")])
 def test_quantize_nbits_appends_q_suffix(nbits, suffix):
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=False,
        attention_implementation="SPLIT_EINSUM",
        quantize_nbits=nbits,
    )
    assert out == f"dreamshaper_8_1x512x512_se{suffix}"
 def test_quantize_nbits_with_controlnet_and_lora():
    out = compose_out_name(
        ckpt_name="dreamshaper_8.safetensors",
        batch_size=1, width=512, height=512,
        controlnet_support=True,
        attention_implementation="SPLIT_EINSUM",
        lora_names=["a.safetensors"],
        quantize_nbits="6",
    )
    assert out == "dreamshaper_8_a_1x512x512_cn_se_q6"
 def test_quantize_nbits_invalid_raises():
    import pytest as _pytest
    with _pytest.raises(ValueError, match="quantize_nbits"):
        compose_out_name(
            ckpt_name="x.safetensors",
            batch_size=1, width=512, height=512,
            controlnet_support=False,
            attention_implementation="SPLIT_EINSUM",
            quantize_nbits="16",  # not in {none, 8, 6, 4}
        )
@@ -1,183 +0,0 @@
 from types import SimpleNamespace
 import torch
 from coreml_suite.conversion.attention import (
    SplitEinsumAttnProcessor,
    SplitEinsumV2AttnProcessor,
    apply_attention_implementation,
    split_einsum,
    split_einsum_v2,
 )
 from coreml_suite.conversion.shapes import conv2d_output_shape
 from coreml_suite.conversion.unet import CoreMLUNetWrapper
 class RecordingUNet(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.call = None
    def forward(
        self,
        sample,
        timestep,
        encoder_hidden_states,
        timestep_cond=None,
        added_cond_kwargs=None,
        down_block_additional_residuals=None,
        mid_block_additional_residual=None,
        return_dict=True,
        **kwargs,
    ):
        self.call = {
            "sample": sample,
            "timestep": timestep,
            "encoder_hidden_states": encoder_hidden_states,
            "timestep_cond": timestep_cond,
            "added_cond_kwargs": added_cond_kwargs,
            "down_block_additional_residuals": down_block_additional_residuals,
            "mid_block_additional_residual": mid_block_additional_residual,
            "return_dict": return_dict,
        }
        return (sample + 1,)
 def test_conv2d_output_shape_matches_torch_conv2d_contract():
    conv = torch.nn.Conv2d(
        4,
        8,
        kernel_size=(3, 5),
        stride=(2, 3),
        padding=(1, 2),
        dilation=(1, 2),
    )
    assert conv2d_output_shape(17, 19, conv) == (9, 5)
 def test_unet_wrapper_passes_context_through_for_sd15():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SD15"))
    sample = torch.randn(2, 4, 8, 8)
    timestep = torch.randn(2)
    context = torch.randn(2, 77, 768)
    out = wrapper(sample, timestep, context)
    assert torch.equal(out, sample + 1)
    assert unet.call["encoder_hidden_states"] is context
    assert unet.call["return_dict"] is False
 def test_unet_wrapper_routes_lcm_sdxl_and_controlnet_inputs():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="LCM"))
    sample = torch.randn(1, 4, 8, 8)
    timestep = torch.randn(1)
    context = torch.randn(1, 77, 768)
    timestep_cond = torch.randn(1, 256)
    down_residual = torch.randn(1, 320, 8, 8)
    mid_residual = torch.randn(1, 1280, 1, 1)
    wrapper(sample, timestep, context, timestep_cond, down_residual, mid_residual)
    assert unet.call["timestep_cond"] is timestep_cond
    assert len(unet.call["down_block_additional_residuals"]) == 1
    assert unet.call["down_block_additional_residuals"][0] is down_residual
    assert unet.call["mid_block_additional_residual"] is mid_residual
 def test_unet_wrapper_routes_sdxl_added_conditioning():
    unet = RecordingUNet()
    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SDXL"))
    sample = torch.randn(1, 4, 8, 8)
    timestep = torch.randn(1)
    context = torch.randn(1, 77, 2048)
    time_ids = torch.randn(1, 6)
    text_embeds = torch.randn(1, 1280)
    wrapper(sample, timestep, context, time_ids, text_embeds)
    assert unet.call["added_cond_kwargs"]["time_ids"] is time_ids
    assert unet.call["added_cond_kwargs"]["text_embeds"] is text_embeds
 def test_split_einsum_matches_original_attention_math():
    torch.manual_seed(0)
    batch = 2
    heads = 3
    dim_head = 4
    sequence = 16
    channels = heads * dim_head
    q = torch.randn(batch, channels, 1, sequence)
    k = torch.randn(batch, channels, 1, sequence)
    v = torch.randn(batch, channels, 1, sequence)
    expected = _original_attention(q, k, v, None, heads, dim_head)
    # split-einsum reorders the float32 reductions vs the reference, so equality
    # only holds up to rounding; the drift exceeds allclose's default atol on
    # some BLAS backends (e.g. Linux x86 CI).
    assert torch.allclose(split_einsum(q, k, v, None, heads, dim_head), expected, atol=1e-6)
    assert torch.allclose(split_einsum_v2(q, k, v, None, heads, dim_head), expected, atol=1e-6)
 def test_split_einsum_v2_chunked_path_matches_original_attention_math():
    torch.manual_seed(0)
    batch = 1
    heads = 2
    dim_head = 2
    sequence = 512
    channels = heads * dim_head
    q = torch.randn(batch, channels, 1, sequence)
    k = torch.randn(batch, channels, 1, sequence)
    v = torch.randn(batch, channels, 1, sequence)
    expected = _original_attention(q, k, v, None, heads, dim_head)
    assert torch.allclose(
        split_einsum_v2(q, k, v, None, heads, dim_head),
        expected,
        atol=1e-6,
    )
 def test_apply_attention_implementation_sets_split_processors():
    unet = RecordingProcessorUNet()
    assert apply_attention_implementation(unet, "ORIGINAL") is unet
    assert unet.processor is None
    apply_attention_implementation(unet, "SPLIT_EINSUM")
    assert isinstance(unet.processor, SplitEinsumAttnProcessor)
    apply_attention_implementation(unet, "SPLIT_EINSUM_V2")
    assert isinstance(unet.processor, SplitEinsumV2AttnProcessor)
 class RecordingProcessorUNet:
    def __init__(self):
        self.processor = None
    def set_attn_processor(self, processor):
        self.processor = processor
 def _original_attention(q, k, v, mask, heads, dim_head):
    batch = q.size(0)
    mh_q = q.view(batch, heads, dim_head, -1)
    mh_k = k.view(batch, heads, dim_head, -1)
    mh_v = v.view(batch, heads, dim_head, -1)
    weights = torch.einsum("bhcq,bhck->bhqk", mh_q, mh_k)
    weights = weights * (dim_head**-0.5)
    if mask is not None:
        weights = weights + mask
    weights = weights.softmax(dim=3)
    attn = torch.einsum("bhqk,bhck->bhcq", weights, mh_v)
    return attn.contiguous().view(batch, heads * dim_head, 1, -1)
Author	SHA1	Message	Date
aszc-dev	1008f144ad	docs: sync with converter consolidation Docs described the pre-#67 suite: a removed standalone LCM converter node, a removed model_version input, a nonexistent CoreMLDetailerHookProvider node, and a deleted lcm/converter.py file. - drop LCM converter node docs; LCM checkpoints are auto-detected by the consolidated CoreMLConverter - remove model_version input and invented 512-768 resolution range - replace phantom detailer-hook fix with real workarounds - fix Python support claim (3.12+ per requires-python) - update LCM support-matrix row, drop stale line reference Refs #67, #68	2026-07-09 18:33:17 +02:00
aszc-dev	8f94f0eea5	docs: rewrite README and split into docs/ pages Rewrite the README as a lean landing page and move depth into a docs/ folder. Correct the supported-model story and several stale facts, and answer the recurring questions from issue #21. - Convert-only is the supported path: suite-converted .mlpackage is the only supported input; drop coreml-community download guidance. - Remove all .mlmodelc / Xcode references — compilation was dropped and the loader handles .mlpackage only. - Fix compute-unit name (CPU_AND_NE, not CPU_AND_ANE) and the loader input name (coreml_name). - Document CoreMLSamplerAdvanced (previously undocumented). - Add docs/: hardware, nodes, conversion, workflows, faq, troubleshooting, limitations (with a support matrix). - Note conversion now lives in the coreml-diffusion package. - Remove dev scaffolding specs; ignore *.log, .DS_Store, .claude/.	2026-07-09 18:30:26 +02:00
aszc	5e57024336	chore(deps): drop <3.13 Python cap, require coreml-diffusion>=0.1.6 (#68 ) Mirror the library: coreml-diffusion 0.1.6 dropped its stale <3.13 cap, so this package can too (requires-python >=3.12). Bump the dependency floor to 0.1.6 to keep the cap consistent across both — a consumer pinned to <3.13 with a library allowing 3.13+ would be unresolvable for the 3.13 range. Pin the dev/CI interpreter to 3.12 via .python-version so CI runs on a fixed baseline rather than floating to a newer Python without a pinned-torch wheel.	2026-06-13 15:48:42 +02:00
aszc	8d28964831	feat(convert): consolidate into one auto-detecting converter node (#67 ) * feat(lcm): convert any full-distill LCM checkpoint - COREML_CONVERT_LCM gains a ckpt_name input: any checkpoint from the checkpoints folder, with the canonical SimianLuo single file as the default auto-download entry, so workflows saved before this input existed keep the old behavior - conversion routes through the unified coreml_diffusion.convert(model_version=LCM) path; the bespoke trace/convert pipeline in lcm/converter.py and the dead UNet2DConditionModelLCM wrapper are removed - output naming via compose_out_name; existing cached LCM .mlpackages reconvert once due to the new _se attention suffix in the name - LCM-LoRA merged checkpoints (plain SD1.5 architecture, no guidance embedding) are rejected by the package with a pointer to the standard converter node + LCM scheduler - requires coreml-diffusion>=0.1.4 (generic LCM conversion fix) Verified against a local ComfyUI checkout: the default entry resolves the Hugging Face single file and cache-hits a previously converted .mlpackage exposing timestep_cond; an LCM-LoRA merge raises the explanatory ValueError. * feat(convert): consolidate conversion into one auto-detecting node The standard CoreMLConverter now auto-detects the model version from the checkpoint (coreml-diffusion>=0.1.5, convert(model_version=None)), so: - the model_version dropdown is gone — one node converts SD15 / SDXL / SDXL refiner / full-distill LCM, the version inferred from the UNet architecture - the dedicated "Core ML LCM Converter" node, its single-model autodownload, and coreml_suite/lcm/nodes.py are removed; the converter UX was previously inconsistent (LCM only reachable through a separate autodownload-only node, while the standard converter did not list LCM at all) lcm/utils.py (sampler-side timestep_cond patching) is unchanged — runtime LCM support still keys off the converted UNet exposing timestep_cond. diffusers is dropped from the dependencies (no longer imported directly after the LCM converter removal). The e2e workflow fixture drops its now-invalid model_version input. * fix(deps): require coreml-diffusion>=0.1.5, keep requires-python <3.13 The auto-detect consolidation needs convert(model_version=None) from coreml-diffusion 0.1.5. requires-python stays pinned to <3.13 to match the library (coremltools-driven); relaxing it past the library's own cap makes the dependency unresolvable for the 3.13+ range.	2026-06-13 14:11:20 +02:00
aszc	a199e749bc	fix(deps): unpin torch; require coreml-diffusion>=0.1.1 (#65 ) The transitive torch<2.8 cap (lifted in coreml-diffusion 0.1.1) downgraded the host's torch on install and broke ComfyUI startup (operator torchvision::nms does not exist). Drop the redundant direct torch pin; torch is provided by the host and pulled transitively as >=2.7.	2026-05-27 05:56:45 +02:00
aszc	a3d555cdc5	fix(registry): depend on PyPI coreml-diffusion, set icon, fix license (#64 ) * fix(registry): depend on PyPI coreml-diffusion, set icon, fix license The git+ dependency on coreml-diffusion tripped the registry security scan (versions flagged unsafe -> Latest pinned to 1.0.1) and broke the node-extraction sandbox (import failed -> 'No nodes found'). - depend on coreml-diffusion from PyPI (>=0.1.0,<0.2); drop git source - set the registry Icon to the bundled 512x512 snake.png - relicense LICENSE file to MIT to match pyproject (was stray GPLv3) - bump version to 2.1.1 to trigger a fresh registry publish Note: uv.lock still references the old git source; refresh with 'uv lock' once coreml-diffusion 0.1.0 is live on PyPI. * chore: refresh lockfile against PyPI coreml-diffusion	2026-05-27 02:54:12 +02:00
aszc	d90546b6bb	feat(extraction): split conversion into the coreml-diffusion package (E0–E5) (#63 ) * docs(extraction): E0 seam inventory + correct stale spec assumptions Resolve all pre-flight greps for the converter-extraction seam: - conversion/* and lcm/unet.py confirmed comfy-free - converter.py: only folder_paths reach-in is get_out_path - lcm/converter.py: folder_paths + comfy.model_management to cut; dup helpers and SimianLuo HF hardcode confirmed - no attention module-global; already per-call Correct two stale assumptions verified against current source: - ml-stable-diffusion is already fully removed (#58); CoreMLModel is a local coremltools wrapper, not Apple's. Drop the dep-pinning blocker and the package/suite dep lines that assumed it. - model_version discovery must emit .name (node reverses via ModelVersion[...]); the .value form in the draft would KeyError on every saved workflow. * feat(extraction): E1 coreml_diffusion package + discovery API Stand up the framework-free coreml_diffusion namespace and freeze its versioned discovery contract. The package re-exports from already comfy-free coreml_suite sources (model_version, attention, core.naming); the conversion implementation moves in E2. - list_model_versions/list_attention_impls/list_quant_modes return today's exact dropdown strings, so wiring the node onto them (E3) changes no value and breaks no saved workflow. - Status/_MODEL_STATUS registry gates VERIFIED vs EXPERIMENTAL in the package, so promoting a model expands the node dropdown with no Suite change (additive-only contract; CONTRACT_VERSION=1.0). - Tier-0 test pins the contract and proves comfy/diffusers/coremltools are not pulled on import. Node untouched; zero behavior change. * refactor(extraction): E2 move conversion mechanics into coreml_diffusion Physically relocate the framework-free conversion code into the package and collapse the duplicated LCM/main helpers, behavior-preserving. - coreml_suite/conversion/ -> coreml_diffusion/conversion/ (attention, shapes, trace, unet) - coreml_suite/core/naming.py -> coreml_diffusion/naming.py (the cache-key contract now lives with the package; tests re-pointed) - coreml_suite/converter.py logic -> coreml_diffusion/convert.py, with convert() made keyword-only past (ckpt_path, model_version, out_path) per the interface contract; out_path is injected (no folder_paths) - dedup: load_coreml_model / convert_to_coreml / get_coreml_inputs / add_cnet_support / get_encoder_hidden_states_shape / inputs-spec now defined once in the package; get_sample_input gains an optional scheduler arg so the LCM path shares it (same keys/order/dtypes) - coreml_suite.{converter,lcm.converter} reduced to comfy-side shims: folder_paths path resolution and the LCM scheduler's comfy.model_management stay here; the package imports neither - __init__ keeps discovery + compose_out_name eager; convert is lazy via __getattr__ so 'import coreml_diffusion' stays Tier-0 pure Nodes untouched (E3 thins them onto the package). Tier-0 (109) and smoke (3, real coremltools conversion) green; [M2-ANE] golden pending a server. * refactor(extraction): E3 thin nodes onto coreml_diffusion + discovery dropdowns The CoreMLConverter node now calls coreml_diffusion directly instead of the coreml_suite.converter shim, and its dropdowns are populated at runtime from the package's discovery API. - INPUT_TYPES dropdowns (model_version / attention_implementation / quantize_nbits) now come from a fail-soft _discover() that calls coreml_diffusion.list_; a missing/old package falls back to a literal list and logs a warning instead of de-registering the node. Installing a newer coreml_diffusion surfaces new conversion types with no Suite change. - folder_paths path resolution moved inline into the node; the package's convert() takes the output path as an injected positional. - compose_out_name / lora_names_from_params now imported from coreml_diffusion (lazily, inside convert) — no node-side copy. - deleted the dead coreml_suite/converter.py and coreml_suite/core/naming.py shims (no remaining importers). Field names, RETURN_TYPES/NAMES and NODE__MAPPINGS unchanged; dropdown values are a superset of the prior literals (additive-only). Tier-0 (109) and smoke (3) green; [M2-ANE] golden re-runs on push. * refactor(extraction): E5 depend on external coreml-diffusion package Conversion code now lives in the standalone coreml-diffusion repo. The Suite deletes its in-tree copy and depends on the package instead. - removed coreml_diffusion/ (whole package), coreml_suite/model_version.py and coreml_suite/attention.py (moved to the package as its source of truth), and the tests that moved with them (discovery, conversion_helpers, out_name; smoke synthetic_unet + split_einsum) - re-pointed ModelVersion imports (config.py, nodes.py, lcm/converter.py) to coreml_diffusion - pyproject: drop the coreml_diffusion package include and the conversion-only deps (peft/omegaconf/transformers, now transitive via coreml-diffusion); add coreml-diffusion as a dependency with a local path source until it is published (switch to git tag/PyPI once the repo exists, so CI can resolve it) Suite Tier-0 green (75); conversion code fully absent from the Suite. The comfy node still imports coreml_diffusion (installed package) for ModelVersion + the discovery dropdowns + convert. * build(extraction): pin coreml-diffusion to git tag v0.1.0 Switch the coreml-diffusion source from a local path to the published git tag so CI can resolve it. Suite Tier-0 green resolving from the tag. * ci(extraction): drop Suite smoke tier (moved to coreml-diffusion) The conversion smoke tests moved to the coreml-diffusion repo, which runs its own Tier 1. The Suite's smoke lane had no tests left (pytest exit 5). The Suite keeps Tier 0 (inference units) and the m2 golden e2e. * chore(release): v2.1.0; wire coreml-diffusion into requirements.txt Minor bump: the conversion path moved to the external coreml-diffusion package (node graph + artifact cache keys unchanged, golden-verified). requirements.txt (used by ComfyUI Manager) now installs coreml-diffusion from the v0.1.0 tag and drops the conversion-only deps now provided transitively.	2026-05-26 22:12:56 +02:00
aszc-dev	f054bbe991	docs: update Stable Diffusion 1.5 links	2026-05-26 22:01:31 +02:00