docs: clarify PYTORCH_ENABLE_MPS_FALLBACK FAQ entry

Added overrides decorator

.github/workflows/publish.yml

@@ -1,25 +0,0 @@
-name: Publish to Comfy registry
-on:
-  workflow_dispatch:
-  push:
-    branches:
-      - main
-    paths:
-      - "pyproject.toml"
-
-permissions:
-  issues: write
-
-jobs:
-  publish-node:
-    name: Publish Custom Node to registry
-    runs-on: ubuntu-latest
-    if: ${{ github.repository_owner == 'aszc-dev' }}
-    steps:
-      - name: Check out code
-        uses: actions/checkout@v4
-      - name: Publish Custom Node
-        uses: Comfy-Org/publish-node-action@v1
-        with:
-          ## Add your own personal access token to your Github Repository secrets and reference it here.
-          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.github/workflows/tier0.yml

@@ -1,27 +0,0 @@
-name: Tier 0 — Unit (Linux)
-
-on:
-  push:
-    branches: [main]
-  pull_request:
-
-# Deps are resolved from pyproject.toml via uv, so the toolchain pins live in
-# one place. Tier 0 must run without ComfyUI; the in-tree purity gate
-# (tests/unit/test_tier0_purity.py) enforces that the suite hasn't started
-# leaking framework imports.
-jobs:
-  unit:
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: astral-sh/setup-uv@v7
-        with:
-          enable-cache: true
-
-      - name: uv sync
-        run: uv sync --no-install-project
-
-      - name: Run Tier 0
-        run: uv run pytest -m unit tests/ -v

.github/workflows/tier2.yml

@@ -1,134 +0,0 @@
-name: Tier 2 — M2 / ANE (self-hosted)
-
-on:
-  pull_request:
-    # `labeled` fires when run-m2 is first added; `synchronize`/`reopened`
-    # re-run on every subsequent push while the label is present, so the
-    # result tracks the PR head instead of going stale. The `if` below keeps
-    # the run gated on the run-m2 label for all pull_request events.
-    types: [labeled, synchronize, reopened]
-  schedule:
-    # Nightly at 04:00 UTC (~05/06 in PL). Keeps the M2 path honest
-    # without burning the runner on every PR.
-    - cron: "0 4 * * *"
-  workflow_dispatch:
-
-jobs:
-  m2:
-    if: |
-      github.event_name == 'schedule' ||
-      github.event_name == 'workflow_dispatch' ||
-      (github.event_name == 'pull_request' &&
-       contains(github.event.pull_request.labels.*.name, 'run-m2'))
-    # Self-hosted Apple Silicon runner. Prerequisites: COMFY_DIR pointing at
-    # a runner-owned ComfyUI clone, plus a cached SD1.5 checkpoint.
-    runs-on: [self-hosted, macOS, ARM64, coreml]
-    timeout-minutes: 90
-    steps:
-      - uses: actions/checkout@v4
-
-      # Hybrid ComfyUI strategy:
-      #   - schedule (nightly)  -> latest origin/master + ComfyUI's own
-      #     requirements.txt (constrained). Canary for upstream API breakage.
-      #   - PR label / dispatch -> the requires-comfyui version tag + the frozen
-      #     `comfy` uv group. Reproducible merge gate, immune to overnight drift.
-      - name: Resolve ComfyUI ref + mode
-        run: |
-          if [ "$GITHUB_EVENT_NAME" = "schedule" ]; then
-            echo "COMFY_MODE=latest" >> "$GITHUB_ENV"
-            echo "COMFY_REF=master"  >> "$GITHUB_ENV"
-          else
-            # requires-comfyui is a semver constraint (e.g. ">=0.3.27"); pin the
-            # gate to the matching ComfyUI release tag (vX.Y.Z).
-            VERSION="$(sed -nE 's/^requires-comfyui *= *"[^0-9]*([0-9]+\.[0-9]+\.[0-9]+).*/\1/p' pyproject.toml)"
-            if [ -z "$VERSION" ]; then echo "could not parse requires-comfyui from pyproject.toml"; exit 1; fi
-            echo "COMFY_MODE=pinned"      >> "$GITHUB_ENV"
-            echo "COMFY_REF=v$VERSION"    >> "$GITHUB_ENV"
-          fi
-
-      - name: Set up ComfyUI checkout
-        # COMFY_DIR is exported by the self-hosted runner's .env and MUST be a
-        # runner-owned ComfyUI clone (never your dev checkout — this step does
-        # git reset --hard and rewrites custom_nodes). Cloned on first run.
-        run: |
-          set -euo pipefail
-          if [ -z "${COMFY_DIR:-}" ]; then echo "COMFY_DIR unset"; exit 1; fi
-          # Init-in-place rather than `git clone`: COMFY_DIR may already hold the
-          # cached checkpoint (models/checkpoints) or converted .mlmodelc, and
-          # `git clone` refuses a non-empty target. init + fetch + `checkout -f`
-          # populates the ComfyUI tree while leaving untracked files (the
-          # checkpoint, the cached models) untouched — so setup order is free.
-          if [ ! -d "$COMFY_DIR/.git" ]; then
-            echo "initialising ComfyUI repo in $COMFY_DIR"
-            mkdir -p "$COMFY_DIR"
-            git -C "$COMFY_DIR" init -q
-          fi
-          git -C "$COMFY_DIR" remote get-url origin >/dev/null 2>&1 \
-            || git -C "$COMFY_DIR" remote add origin https://github.com/comfyanonymous/ComfyUI.git
-          git -C "$COMFY_DIR" fetch --quiet origin
-          if [ "$COMFY_MODE" = "latest" ]; then
-            git -C "$COMFY_DIR" checkout -f -B master origin/master
-          else
-            git -C "$COMFY_DIR" checkout -f "$COMFY_REF"
-          fi
-          COMFY_SHA="$(git -C "$COMFY_DIR" rev-parse HEAD)"
-          echo "COMFY_SHA=$COMFY_SHA" >> "$GITHUB_ENV"
-          echo "Tier 2 mode=$COMFY_MODE, ComfyUI \`$COMFY_SHA\`" >> "$GITHUB_STEP_SUMMARY"
-
-          # Point ComfyUI's custom-node loader at this checkout. Refresh the
-          # symlink only; refuse to clobber a real directory (guards against a
-          # COMFY_DIR that is accidentally a dev checkout).
-          NODE_LINK="$COMFY_DIR/custom_nodes/ComfyUI-CoreMLSuite"
-          if [ -e "$NODE_LINK" ] && [ ! -L "$NODE_LINK" ]; then
-            echo "ERROR: $NODE_LINK is a real directory, not a symlink."
-            echo "COMFY_DIR must be a runner-owned ComfyUI, not your dev checkout."
-            exit 1
-          fi
-          mkdir -p "$COMFY_DIR/custom_nodes"
-          ln -sfn "$GITHUB_WORKSPACE" "$NODE_LINK"
-
-      - name: Install dependencies
-        run: |
-          set -euo pipefail
-          if [ "$COMFY_MODE" = "latest" ]; then
-            # Node deps (our coremltools-9 toolchain), then ComfyUI's own
-            # requirements for the pulled SHA, capped by the toolchain ceiling.
-            uv sync
-            uv pip install -r "$COMFY_DIR/requirements.txt" \
-              -c constraints/comfy-ceiling.txt
-          else
-            # Pinned gate: the frozen group mirrors the known-good pinned SHA.
-            uv sync --group comfy
-          fi
-
-      - name: Start ComfyUI server (background)
-        run: |
-          cd "$COMFY_DIR"
-          nohup "$GITHUB_WORKSPACE/.venv/bin/python" main.py --port 8188 --cpu-vae > /tmp/comfyui-ci.log 2>&1 &
-          # Poll the HTTP endpoint for readiness — robust to startup-banner
-          # wording / colored-log changes in a floating-latest ComfyUI.
-          for _ in $(seq 1 90); do
-            if curl -sf -o /dev/null http://127.0.0.1:8188/system_stats; then
-              echo "comfy ready (ComfyUI ${COMFY_SHA:-unknown})"; exit 0
-            fi
-            sleep 2
-          done
-          echo "comfy failed to start"; tail -100 /tmp/comfyui-ci.log; exit 1
-
-      - name: Purge cached Core ML UNets (force fresh conversion)
-        # The converter skips when a model of the same name already exists. That
-        # cache key is conversion *parameters* only, not the conversion code or
-        # toolchain — so a stale model would let a conversion regression pass.
-        # Clear it so every Tier 2 run exercises the full convert -> compile ->
-        # sample path end to end.
-        run: |
-          rm -rf "$COMFY_DIR"/models/unet/*.mlpackage "$COMFY_DIR"/models/unet/*.mlmodelc || true
-
-      - name: Run Tier 2 (m2 marker)
-        # Drives the Core ML Converter node, which converts the UNet from the
-        # checkpoint on every run (cache purged above).
-        run: uv run --no-sync pytest -m m2 tests/ -v
-
-      - name: Stop ComfyUI server
-        if: always()
-        run: pkill -f "main.py.*8188" || true

.gitignore

@@ -1,8 +1,3 @@
 playground/
+experiments/
 __pycache__/
-models/
-.venv/
-test_results/
-*.log
-.DS_Store
-.claude/

.python-version

@@ -1 +0,0 @@
-3.12

LICENSE

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

README.md

@@ -1,113 +1,432 @@
 # Core ML Suite for ComfyUI
 
-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.
+## Overview
 
-You convert a Stable Diffusion checkpoint to a Core ML model with the nodes in
-this suite, then sample from it like any other ComfyUI workflow.
+Welcome! In this repository you'll find a set of custom nodes for [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
+that allows you to use Core ML models in your ComfyUI workflows.
+These models are designed to leverage the Apple Neural Engine (ANE) on Apple Silicon (M1/M2) machines,
+thereby enhancing your workflows and improving performance.
 
-> [!IMPORTANT]
-> **Convert your own checkpoints — that is the only supported path.** This
-> 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.
+If you're not sure how to obtain these models, you can download them
+[here](https://huggingface.co/coreml-community) or convert your own models using
+[coremltools](https://github.com/apple/ml-stable-diffusion).
 
-## Installation
+In simple terms, think of Core ML models as a tool that can help your ComfyUI work faster and more efficiently.
+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.
 
-### ComfyUI-Manager (recommended)
+## Getting Started
 
-Open **Manager → Install Custom Nodes**, search for `Core ML`, click
-**Install**, and restart ComfyUI.
+To start using custom nodes in your ComfyUI, follow these simple steps:
 
-### Manual
+1. Clone or download this repository: You can do this directly into the custom_nodes directory of your ComfyUI.
+2. Install the dependencies: You'll need to use a package manager like pip to do this.
 
-```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
-```
+That's it! You're now ready to start enhancing your ComfyUI workflows with Core ML models.
 
-Dependencies (`coreml-diffusion`, `coremltools`, `numpy`, `diffusers`) install
-from PyPI. PyTorch is intentionally **not** pinned — it is provided by your
-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.
+- Check [Installation](#installation) for more details on installation.
+- Check [How to use](#how-to-use) for more details on how to use the custom nodes.
+- Check [Example Workflows](#example-workflows) for some example workflows.
 
 ## Glossary
 
-- **Core ML** — Apple's on-device machine-learning framework.
-- **`.mlpackage`** — the Core ML model format this suite produces and loads.
-- **ANE** — Apple Neural Engine, a hardware accelerator for ML.
-- **Compute unit** — which hardware Core ML uses (`CPU_AND_NE`, `CPU_AND_GPU`,
-  `CPU_ONLY`, `ALL`).
-- **Attention implementation** — `SPLIT_EINSUM` / `SPLIT_EINSUM_V2` (ANE-friendly)
-  or `ORIGINAL` (GPU-friendly), chosen at conversion.
+- **Core ML**: A machine learning framework developed by Apple. It's used to run machine learning models on Apple
+  devices.
+- **Core ML Model**: A machine learning model that can be run on Apple devices using Core ML.
+- **mlmodelc**: A compiled Core ML model. This is the recommended format for Core ML models.
+- **mlpackage**: A Core ML model packaged in a directory. This is the default format for Core ML models.
+- **ANE**: Apple Neural Engine. A hardware accelerator for machine learning tasks on Apple devices.
+- **Compute Unit**: A Core ML option that allows you to specify the hardware on which the model should run.
+    - **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.
+
+## Installation
+
+### Using ComfyUI-Manager
+
+The easiest way to install the custom nodes is to use the ComfyUI-Manager. You can find the installation instructions
+[here](https://github.com/ltdrdata/ComfyUI-Manager#installation). Once you've installed the ComfyUI-Manager, you can
+install the custom nodes by following these steps:
+
+- Open the ComfyUI-Manager by clicking the `Manager` button in the ComfyUI toolbar.
+- Click the `Install Custom Nodes` button.
+- Search for `Core ML` and click the `Install` button.
+- 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. 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.
+    - **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/runwayml/stable-diffusion-v1-5/blob/main/text_encoder/model.safetensors) and
+      [here](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/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/runwayml/stable-diffusion-v1-5/blob/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]
-> **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, hidden, 1, tokens)`.
-> Models converted with earlier versions are not compatible and must be
-> re-converted.
+> 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.
 
-## Acknowledgements
+![sdxl](./assets/sdxl_conversion.png?raw=true)
 
-The conversion pipeline began as an adaptation of Apple's
-[ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion), which
-pioneered running Stable Diffusion on the Neural Engine. It has since diverged
-and no longer depends on that package: UNet conversion runs natively on
-`diffusers`' `UNet2DConditionModel`, the ANE attention path (`SPLIT_EINSUM`,
-`SPLIT_EINSUM_V2`) is reimplemented as standalone `diffusers` attention
-processors, and the toolchain tracks current ComfyUI (NumPy 2, Torch 2.7+,
-coremltools 9, Python 3.12+). Conversion now lives in the separate
-[coreml-diffusion](https://github.com/aszc-dev/coreml-diffusion) package.
+## 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 convert the model to a different input size using tools available
+  in the [apple/ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion) repository.
+- 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.
+
+## FAQ
+
+### Hardware and Performance
+
+#### What's the difference between MPS, GPU, and ANE?
+- **MPS (Metal Performance Shaders)**: Apple's framework for GPU acceleration. It's what PyTorch uses by default on Apple Silicon.
+- **GPU**: The graphics processing unit on your Apple Silicon chip.
+- **ANE (Apple Neural Engine)**: A specialized hardware accelerator for machine learning tasks.
+
+#### Which compute unit should I choose?
+- **CPU_AND_ANE**: Best for models converted with `--attention-implementation SPLIT_EINSUM`. This is the default and recommended option for most users.
+- **CPU_AND_GPU**: Best for models converted with `--attention-implementation ORIGINAL`. Use this if you experience issues with ANE.
+- **CPU_ONLY**: Use this as a fallback if you experience issues with both ANE and GPU.
+
+#### Do I need `PYTORCH_ENABLE_MPS_FALLBACK=1`?
+While our Core ML nodes don't use this environment variable directly, it may still be relevant for other parts of ComfyUI that use PyTorch with MPS backend. The setting of this variable is a user preference and depends on your specific needs and workflow requirements.
+
+### Model Conversion and Compatibility
+
+#### Is there a performance penalty when using the Core ML Adapter?
+Yes, there might be a slight performance penalty compared to using directly converted models. However, the adapter provides more flexibility and compatibility with standard ComfyUI nodes.
+
+#### Does the Core ML Adapter support SDXL?
+Currently, SDXL support in the Core ML Adapter is limited. While it may work with some models, it's not officially supported and may cause issues.
+
+#### Are `mlmodelc` and `mlpackage` formats safe?
+Yes, both formats are safe to use. However, we recommend:
+1. Always downloading original `.safetensors` files from trusted sources
+2. Converting them yourself using our tools
+3. Using the converted `.mlmodelc` files for better performance
+
+#### Do Core ML models produce identical results to their safetensors counterparts?
+While the results should be very similar, there might be slight differences due to:
+- Different numerical precision
+- Hardware-specific optimizations
+- Different attention implementations
+
+#### Should I convert models every time I queue a generation?
+No! The conversion only happens once when you first use the converter node. After that, you should use the `CoreMLUnetLoader` to load the already converted model.
+
+#### Will SDXL ever be supported on ANE?
+Currently, there are technical limitations preventing SDXL from running efficiently on ANE. We recommend using `CPU_AND_GPU` or `CPU_ONLY` for SDXL models.
 
 ## Support
 
-Questions or suggestions? Open an
-[issue](https://github.com/aszc-dev/ComfyUI-CoreMLSuite/issues).
+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
+to assist you.

__init__.py

@@ -11,6 +11,9 @@ from coreml_suite.nodes import (
     CoreMLConverter,
     COREML_LOAD_LORA,
 )
+from coreml_suite.lcm import (
+    COREML_CONVERT_LCM,
+)
 
 NODE_CLASS_MAPPINGS = {
     "CoreMLUNetLoader": CoreMLLoaderUNet,
@@ -19,6 +22,7 @@ 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",
@@ -27,4 +31,5 @@ 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",
 }

conftest.py

@@ -1,4 +0,0 @@
-"""Top-level conftest: prevent pytest from importing the repo-root
-__init__.py (the ComfyUI custom-node entry point pulls in comfy + nodes,
-which breaks the Tier-0 'no-framework' promise)."""
-collect_ignore = ["__init__.py"]

constraints/comfy-ceiling.txt

@@ -1,18 +0,0 @@
-# Toolchain ceiling for installing a floating-latest ComfyUI's requirements.txt
-# in the Tier 2 nightly canary (.github/workflows/tier2.yml, latest mode).
-#
-# ComfyUI's requirements.txt requests bare `torch`/`torchvision`/`torchaudio`
-# and `numpy>=1.25.0`, which would float past the versions coremltools 9 /
-# apple-ml-stable-diffusion have been validated against.
-# These constraints cap the resolution so the canary keeps testing the same
-# toolchain the suite actually ships.
-#
-# If upstream ComfyUI ever hard-requires something beyond these bounds, the
-# install FAILS — and that failure is the signal we want: it means the host
-# outgrew the pinned toolchain and coremltools / ml-stable-diffusion need a
-# deliberate bump, not a silent float.
-torch>=2.7,<2.8
-torchvision>=0.22,<0.23
-torchaudio>=2.7,<2.8
-numpy>=1.25,<2
-coremltools>=9,<10

coreml_suite/config.py

@@ -1,10 +1,17 @@
+from enum import Enum
+
 import torch
 
 from comfy import supported_models_base
 from comfy import latent_formats
 from comfy.model_detection import convert_config
 
-from coreml_diffusion import ModelVersion
+
+class ModelVersion(Enum):
+    SD15 = "sd15"
+    SDXL = "sdxl"
+    SDXL_REFINER = "sdxl_refiner"
+    LCM = "lcm"
 
 
 config_map = {

coreml_suite/controlnet.py

@@ -1,14 +1,62 @@
-"""Compatibility shim — re-exports from coreml_suite.core.controlnet."""
-from coreml_suite.core.controlnet import (
-    chunk_control,
-    expand_inputs,
-    extract_residual_kwargs,
-    no_control,
-)
+from itertools import chain
+from math import ceil
 
-__all__ = [
-    "chunk_control",
-    "expand_inputs",
-    "extract_residual_kwargs",
-    "no_control",
-]
+import numpy as np
+import torch
+
+from coreml_suite.latents import chunk_batch
+
+
+def expand_inputs(inputs):
+    expanded = inputs.copy()
+    for k, v in inputs.items():
+        if isinstance(v, np.ndarray):
+            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, torch.Tensor):
+            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, list):
+            expanded[k] = v * 2 if len(v) == 1 else v
+        elif isinstance(v, dict):
+            expand_inputs(v)
+    return expanded
+
+
+def extract_residual_kwargs(expected_inputs, control):
+    if "additional_residual_0" not in expected_inputs.keys():
+        return {}
+    if control is None:
+        return no_control(expected_inputs)
+
+    residual_kwargs = {
+        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
+        for i, r in enumerate(chain(control["output"], control["middle"]))
+    }
+    return residual_kwargs
+
+
+def no_control(expected_inputs):
+    shapes_dict = {
+        k: v["shape"] for k, v in expected_inputs.items() if k.startswith("additional")
+    }
+    residual_kwargs = {
+        k: torch.zeros(*shape).cpu().numpy().astype(dtype=np.float16)
+        for k, shape in shapes_dict.items()
+    }
+    return residual_kwargs
+
+
+def chunk_control(cn, target_size):
+    if cn is None:
+        return [None] * target_size
+
+    num_chunks = ceil(cn["output"][0].shape[0] / target_size)
+
+    out = [{"output": [], "middle": []} for _ in range(num_chunks)]
+
+    for k, v in cn.items():
+        for i, x in enumerate(v):
+            chunks = chunk_batch(x, (target_size, *x.shape[1:]))
+            for j, chunk in enumerate(chunks):
+                out[j][k].append(chunk)
+
+    return out

coreml_suite/converter.py

@@ -0,0 +1,362 @@
+import gc
+import os
+import shutil
+import time
+from typing import Union
+
+import coremltools as ct
+import numpy as np
+import python_coreml_stable_diffusion.unet
+import torch
+from diffusers import (
+    StableDiffusionPipeline,
+    LatentConsistencyModelPipeline,
+    StableDiffusionXLPipeline,
+)
+from python_coreml_stable_diffusion.unet import (
+    UNet2DConditionModel,
+    UNet2DConditionModelXL,
+    AttentionImplementations,
+)
+
+from coreml_suite.config import ModelVersion
+from coreml_suite.lcm.unet import UNet2DConditionModelLCM
+from coreml_suite.logger import logger
+from folder_paths import get_folder_paths
+
+
+class StableDiffusionLCMPipeline(LatentConsistencyModelPipeline):
+    pass
+
+
+MODEL_TYPE_TO_UNET_CLS = {
+    ModelVersion.SD15: UNet2DConditionModel,
+    ModelVersion.SDXL: UNet2DConditionModelXL,
+    ModelVersion.LCM: UNet2DConditionModelLCM,
+}
+
+MODEL_TYPE_TO_PIPE_CLS = {
+    ModelVersion.SD15: StableDiffusionPipeline,
+    ModelVersion.SDXL: StableDiffusionXLPipeline,
+    ModelVersion.LCM: StableDiffusionLCMPipeline,
+}
+
+
+def get_unet(model_type: ModelVersion, ref_pipe):
+    ref_unet = ref_pipe.unet
+
+    unet_cls = MODEL_TYPE_TO_UNET_CLS[model_type]
+    cml_unet = unet_cls.from_config(ref_unet.config).eval()
+    cml_unet.load_state_dict(ref_unet.state_dict(), strict=False)
+
+    return cml_unet
+
+
+def get_encoder_hidden_states_shape(ref_pipe, batch_size):
+    text_encoder = (
+        ref_pipe.text_encoder_2
+        if hasattr(ref_pipe, "text_encoder_2")
+        else ref_pipe.text_encoder
+    )
+
+    text_token_sequence_length = text_encoder.config.max_position_embeddings
+    hidden_size = (text_encoder.config.hidden_size,)
+
+    encoder_hidden_states_shape = (
+        batch_size,
+        ref_pipe.unet.config.cross_attention_dim or hidden_size,
+        1,
+        text_token_sequence_length,
+    )
+
+    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):
+    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 compile_coreml_model(source_model_path, output_dir, final_name):
+    """Compiles Core ML models using the coremlcompiler utility from Xcode toolchain"""
+    target_path = os.path.join(output_dir, f"{final_name}.mlmodelc")
+    if os.path.exists(target_path):
+        logger.warning(f"Found existing compiled model at {target_path}! Skipping..")
+        return target_path
+
+    logger.info(f"Compiling {source_model_path}")
+    source_model_name = os.path.basename(os.path.splitext(source_model_path)[0])
+
+    os.system(f"xcrun coremlcompiler compile {source_model_path} {output_dir}")
+    compiled_output = os.path.join(output_dir, f"{source_model_name}.mlmodelc")
+    shutil.move(compiled_output, target_path)
+
+    return target_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)),
+        ]
+    )
+    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_pipe):
+    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 = (
+        hasattr(ref_pipe.config, "requires_aesthetics_score")
+        and ref_pipe.config.requires_aesthetics_score
+    )
+
+    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, ref_pipe.text_encoder_2.config.hidden_size)
+
+    return {
+        "time_ids": time_ids,
+        "text_embeds": torch.randn(*text_embeds_shape).to(torch.float32),
+    }
+
+
+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):
+    from python_coreml_stable_diffusion.unet import calculate_conv2d_output_shape
+
+    additional_residuals_shapes = []
+
+    batch_size = sample_shape[0]
+    h, w = sample_shape[2:]
+
+    # conv_in
+    out_h, out_w = calculate_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 = calculate_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_pipe,
+    model_version: ModelVersion,
+    unet_out_path: str,
+    batch_size: int = 1,
+    sample_size: tuple[int, int] = (64, 64),
+    controlnet_support: bool = False,
+):
+    coreml_unet = get_unet(model_version, ref_pipe)
+    ref_unet = ref_pipe.unet
+
+    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_pipe, batch_size)
+
+    scheduler = ref_pipe.scheduler
+    scheduler.set_timesteps(50)
+
+    sample_inputs = get_sample_input(
+        batch_size, encoder_hidden_states_shape, sample_shape, scheduler
+    )
+
+    if model_version == ModelVersion.LCM:
+        sample_inputs |= lcm_inputs(sample_inputs)
+
+    if model_version == ModelVersion.SDXL:
+        sample_inputs |= sdxl_inputs(sample_inputs, ref_pipe)
+
+    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()
+
+    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[Union[str, os.PathLike], float]] = None,
+    attn_impl: str = AttentionImplementations.SPLIT_EINSUM.name,
+    config_path: str = None,
+):
+    if os.path.exists(unet_out_path):
+        logger.info(f"Found existing model at {unet_out_path}! Skipping..")
+        return
+
+    python_coreml_stable_diffusion.unet.ATTENTION_IMPLEMENTATION_IN_EFFECT = (
+        AttentionImplementations(attn_impl)
+    )
+
+    ref_pipe = get_pipeline(ckpt_path, config_path, model_version)
+
+    for i, lora_weight in enumerate(lora_weights or []):
+        lora_path, strength = lora_weight
+        adapter_name = f"lora_{i}"
+        ref_pipe.load_lora_weights(lora_path, adapter_name=adapter_name)
+        ref_pipe.set_adapters([adapter_name], adapter_weights=[strength])
+        ref_pipe.fuse_lora()
+
+    convert_unet(
+        ref_pipe,
+        model_version,
+        unet_out_path,
+        batch_size,
+        sample_size,
+        controlnet_support,
+    )
+
+
+def get_pipeline(ckpt_path, config_path, model_version):
+    pipe_cls = MODEL_TYPE_TO_PIPE_CLS[model_version]
+    ref_pipe = pipe_cls.from_single_file(ckpt_path, original_config_file=config_path)
+    return ref_pipe
+
+
+def compile_model(out_path, out_name, submodule_name):
+    # Compile the model
+    target_path = compile_coreml_model(
+        out_path, get_folder_paths(submodule_name)[0], f"{out_name}_{submodule_name}"
+    )
+    logger.info(f"Compiled {out_path} to {target_path}")
+    return target_path

coreml_suite/core/__init__.py

@@ -1,10 +0,0 @@
-"""Framework-free pure-logic core of ComfyUI-CoreMLSuite.
-
-Modules under this package must NOT import `comfy`, `coremltools`,
-`python_coreml_stable_diffusion`, `folder_paths`, `nodes`, or any other
-ComfyUI / Apple runtime. Only `numpy` and `torch` are allowed.
-
-The thin adapters in `coreml_suite.{latents,controlnet,models}` keep the
-old public import paths working so `coreml_suite/nodes.py` and downstream
-ComfyUI workflows are unchanged.
-"""

coreml_suite/core/controlnet.py

@@ -1,67 +0,0 @@
-"""Pure helpers around the ControlNet residual inputs of the Core ML UNet.
-
-Re-exported by coreml_suite.controlnet. Characterization tests cover
-shapes, dtype (fp16), and zero-fill fallback.
-"""
-from itertools import chain
-from math import ceil
-
-import numpy as np
-import torch
-
-from coreml_suite.core.latents import chunk_batch
-
-
-def expand_inputs(inputs):
-    expanded = inputs.copy()
-    for k, v in inputs.items():
-        if isinstance(v, np.ndarray):
-            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, torch.Tensor):
-            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, list):
-            expanded[k] = v * 2 if len(v) == 1 else v
-        elif isinstance(v, dict):
-            expand_inputs(v)
-    return expanded
-
-
-def extract_residual_kwargs(expected_inputs, control):
-    if "additional_residual_0" not in expected_inputs.keys():
-        return {}
-    if control is None:
-        return no_control(expected_inputs)
-
-    residual_kwargs = {
-        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
-        for i, r in enumerate(chain(control["output"], control["middle"]))
-    }
-    return residual_kwargs
-
-
-def no_control(expected_inputs):
-    shapes_dict = {
-        k: v["shape"] for k, v in expected_inputs.items() if k.startswith("additional")
-    }
-    residual_kwargs = {
-        k: torch.zeros(*shape).cpu().numpy().astype(dtype=np.float16)
-        for k, shape in shapes_dict.items()
-    }
-    return residual_kwargs
-
-
-def chunk_control(cn, target_size):
-    if cn is None:
-        return [None] * target_size
-
-    num_chunks = ceil(cn["output"][0].shape[0] / target_size)
-
-    out = [{"output": [], "middle": []} for _ in range(num_chunks)]
-
-    for k, v in cn.items():
-        for i, x in enumerate(v):
-            chunks = chunk_batch(x, (target_size, *x.shape[1:]))
-            for j, chunk in enumerate(chunks):
-                out[j][k].append(chunk)
-
-    return out

coreml_suite/core/inputs.py

@@ -1,111 +0,0 @@
-"""Pure transform from torch sampler inputs to Core ML UNet kwargs.
-
-Characterization tests cover SD1.5 / SDXL base / SDXL refiner / LCM
-variants and the chunked-batch fan-out.
-"""
-import numpy as np
-import torch
-
-from coreml_suite.core.controlnet import extract_residual_kwargs, chunk_control
-from coreml_suite.core.latents import chunk_batch
-
-
-class CoreMLInputs:
-    def __init__(self, x, t, context, control, **kwargs):
-        self.x = x
-        self.t = t
-        self.context = context
-        self.control = control
-        self.time_ids = kwargs.get("time_ids")
-        self.text_embeds = kwargs.get("text_embeds")
-        self.ts_cond = kwargs.get("timestep_cond")
-
-    def coreml_kwargs(self, expected_inputs):
-        sample = self.x.cpu().numpy().astype(np.float16)
-
-        context = self.context.cpu().numpy().astype(np.float16)
-
-        t = self.t.cpu().numpy().astype(np.float16)
-
-        model_input_kwargs = {
-            "sample": sample,
-            "encoder_hidden_states": context,
-            "timestep": t,
-        }
-        residual_kwargs = extract_residual_kwargs(expected_inputs, self.control)
-        model_input_kwargs |= residual_kwargs
-
-        # LCM
-        if self.ts_cond is not None:
-            model_input_kwargs["timestep_cond"] = (
-                self.ts_cond.cpu().numpy().astype(np.float16)
-            )
-
-        # SDXL
-        if "text_embeds" in expected_inputs:
-            model_input_kwargs["text_embeds"] = (
-                self.text_embeds.cpu().numpy().astype(np.float16)
-            )
-        if "time_ids" in expected_inputs:
-            model_input_kwargs["time_ids"] = (
-                self.time_ids.cpu().numpy().astype(np.float16)
-            )
-
-        return model_input_kwargs
-
-    def chunks(self, expected_inputs):
-        sample_shape = expected_inputs["sample"]["shape"]
-        timestep_shape = expected_inputs["timestep"]["shape"]
-        context_shape = expected_inputs["encoder_hidden_states"]["shape"]
-
-        chunked_x = chunk_batch(self.x, sample_shape)
-        ts = list(torch.full((len(chunked_x), timestep_shape[0]), self.t[0]))
-        chunked_context = chunk_batch(self.context, context_shape)
-
-        chunked_control = [None] * len(chunked_x)
-        if self.control is not None:
-            chunked_control = chunk_control(self.control, sample_shape[0])
-
-        chunked_ts_cond = [None] * len(chunked_x)
-        if self.ts_cond is not None:
-            ts_cond_shape = expected_inputs["timestep_cond"]["shape"]
-            chunked_ts_cond = chunk_batch(self.ts_cond, ts_cond_shape)
-
-        chunked_time_ids = [None] * len(chunked_x)
-        if expected_inputs.get("time_ids") is not None:
-            time_ids_shape = expected_inputs["time_ids"]["shape"]
-            if self.time_ids is None:
-                self.time_ids = torch.zeros(len(chunked_x), *time_ids_shape[1:]).to(
-                    self.x.device
-                )
-            chunked_time_ids = chunk_batch(self.time_ids, time_ids_shape)
-
-        chunked_text_embeds = [None] * len(chunked_x)
-        if expected_inputs.get("text_embeds") is not None:
-            text_embeds_shape = expected_inputs["text_embeds"]["shape"]
-            if self.text_embeds is None:
-                self.text_embeds = torch.zeros(
-                    len(chunked_x), *text_embeds_shape[1:]
-                ).to(self.x.device)
-            chunked_text_embeds = chunk_batch(self.text_embeds, text_embeds_shape)
-
-        return [
-            CoreMLInputs(
-                x,
-                t,
-                context,
-                control,
-                timestep_cond=ts_cond,
-                time_ids=time_ids,
-                text_embeds=text_embeds,
-            )
-            for x, t, context, control, ts_cond, time_ids, text_embeds in zip(
-                chunked_x,
-                ts,
-                chunked_context,
-                chunked_control,
-                chunked_ts_cond,
-                chunked_time_ids,
-                chunked_text_embeds,
-            )
-        ]

coreml_suite/core/latents.py

@@ -1,42 +0,0 @@
-"""Pure batch-chunking helpers for Core ML's fixed-shape UNet inputs.
-
-Re-exported by coreml_suite.latents. Characterization tests cover the
-contract (padding-zero regions, truncation in merge_chunks,
-identity-passthrough when shape already matches).
-"""
-import torch
-
-
-def chunk_batch(input_tensor, target_shape):
-    if input_tensor.shape == target_shape:
-        return [input_tensor]
-
-    batch_size = input_tensor.shape[0]
-    target_batch_size = target_shape[0]
-
-    num_chunks = batch_size // target_batch_size
-    if num_chunks == 0:
-        padding = torch.zeros(target_batch_size - batch_size, *target_shape[1:]).to(
-            input_tensor.device
-        )
-        return [torch.cat((input_tensor, padding), dim=0)]
-
-    mod = batch_size % target_batch_size
-    if mod != 0:
-        chunks = list(torch.chunk(input_tensor[:-mod], num_chunks))
-        padding = torch.zeros(target_batch_size - mod, *target_shape[1:]).to(
-            input_tensor.device
-        )
-        padded = torch.cat((input_tensor[-mod:], padding), dim=0)
-        chunks.append(padded)
-        return chunks
-
-    chunks = list(torch.chunk(input_tensor, num_chunks))
-    return chunks
-
-
-def merge_chunks(chunks, orig_shape):
-    merged = torch.cat(chunks, dim=0)
-    if merged.shape == orig_shape:
-        return merged
-    return merged[: orig_shape[0]]

coreml_suite/core/sdxl.py

@@ -1,91 +0,0 @@
-"""Pure SDXL detection + time_ids/text_embeds assembly.
-
-The framework-coupled adapter `add_sdxl_model_options` lives in models.py
-and delegates the math here. Characterization tests cover base (len 6) vs
-refiner (len 5) and the closure free-vars produced by
-`sdxl_model_function_wrapper`.
-"""
-import torch
-
-
-def is_sdxl(coreml_model):
-    return (
-        "time_ids" in coreml_model.expected_inputs
-        and "text_embeds" in coreml_model.expected_inputs
-    )
-
-
-def is_sdxl_base(coreml_model):
-    return (
-        is_sdxl(coreml_model)
-        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 6
-    )
-
-
-def is_sdxl_refiner(coreml_model):
-    return (
-        is_sdxl(coreml_model)
-        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 5
-    )
-
-
-def build_sdxl_time_ids(pos_dict, neg_dict, *, is_base: bool, is_refiner: bool):
-    """Compose the (2, N) time_ids tensor for the SDXL Core ML UNet.
-
-    - base: N=6  -> [h, w, crop_h, crop_w, target_h, target_w]
-    - refiner: N=5 -> [h, w, crop_h, crop_w, aesthetic_score]
-    - neither: N=4 -> [h, w, crop_h, crop_w] (edge case kept for parity)
-    """
-    pos_time_ids = [
-        pos_dict.get("height", 768),
-        pos_dict.get("width", 768),
-        pos_dict.get("crop_h", 0),
-        pos_dict.get("crop_w", 0),
-    ]
-    neg_time_ids = [
-        neg_dict.get("height", 768),
-        neg_dict.get("width", 768),
-        neg_dict.get("crop_h", 0),
-        neg_dict.get("crop_w", 0),
-    ]
-
-    if is_base:
-        pos_time_ids += [
-            pos_dict.get("target_height", 768),
-            pos_dict.get("target_width", 768),
-        ]
-        neg_time_ids += [
-            neg_dict.get("target_height", 768),
-            neg_dict.get("target_width", 768),
-        ]
-
-    if is_refiner:
-        pos_time_ids += [pos_dict.get("aesthetic_score", 6)]
-        neg_time_ids += [neg_dict.get("aesthetic_score", 2.5)]
-
-    return torch.tensor([pos_time_ids, neg_time_ids])
-
-
-def build_sdxl_text_embeds(pos_pooled, neg_pooled):
-    """Concat pos then neg along the batch dim. Locked contract."""
-    return torch.cat((pos_pooled, neg_pooled))
-
-
-def sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False):
-    def wrapper(model_function, params):
-        x = params["input"]
-        t = params["timestep"]
-        c = params["c"]
-
-        context = c.get("c_crossattn")
-
-        if context is None:
-            return torch.zeros_like(x)
-
-        if refiner and context is not None:
-            # converted refiner accepts only g clip
-            c["c_crossattn"] = context[:, :, 768:]
-
-        return model_function(x, t, **c, time_ids=time_ids, text_embeds=text_embeds)
-
-    return wrapper

coreml_suite/coreml_model.py

@@ -1,42 +0,0 @@
-import time
-
-import coremltools as ct
-
-from coreml_suite.logger import logger
-
-
-class CoreMLModel:
-    """Small runtime wrapper around coremltools.models.MLModel.
-
-    This keeps the inference path independent from apple/ml-stable-diffusion's
-    CoreMLModel wrapper while preserving the contract used by the sampler code:
-    ``expected_inputs`` and callable prediction.
-    """
-
-    def __init__(self, model_path, compute_unit):
-        self.model_path = model_path
-        self.compute_unit = self._compute_unit(compute_unit)
-
-        logger.info(f"Loading {model_path} to {self.compute_unit.name}")
-        start = time.time()
-        self.model = ct.models.MLModel(model_path, compute_units=self.compute_unit)
-        logger.info(f"Loading {model_path} took {time.time() - start:.1f} seconds")
-
-        self.expected_inputs = self._expected_inputs()
-
-    def __call__(self, **kwargs):
-        return self.model.predict(kwargs)
-
-    @staticmethod
-    def _compute_unit(compute_unit):
-        if isinstance(compute_unit, ct.ComputeUnit):
-            return compute_unit
-        return ct.ComputeUnit[compute_unit]
-
-    def _expected_inputs(self):
-        return {
-            feature.name: {
-                "shape": tuple(feature.type.multiArrayType.shape),
-            }
-            for feature in self.model.get_spec().description.input
-        }

coreml_suite/latents.py

@@ -1,4 +1,36 @@
-"""Compatibility shim — re-exports from coreml_suite.core.latents."""
-from coreml_suite.core.latents import chunk_batch, merge_chunks
+import torch
 
-__all__ = ["chunk_batch", "merge_chunks"]
+
+def chunk_batch(input_tensor, target_shape):
+    if input_tensor.shape == target_shape:
+        return [input_tensor]
+
+    batch_size = input_tensor.shape[0]
+    target_batch_size = target_shape[0]
+
+    num_chunks = batch_size // target_batch_size
+    if num_chunks == 0:
+        padding = torch.zeros(target_batch_size - batch_size, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        return [torch.cat((input_tensor, padding), dim=0)]
+
+    mod = batch_size % target_batch_size
+    if mod != 0:
+        chunks = list(torch.chunk(input_tensor[:-mod], num_chunks))
+        padding = torch.zeros(target_batch_size - mod, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        padded = torch.cat((input_tensor[-mod:], padding), dim=0)
+        chunks.append(padded)
+        return chunks
+
+    chunks = list(torch.chunk(input_tensor, num_chunks))
+    return chunks
+
+
+def merge_chunks(chunks, orig_shape):
+    merged = torch.cat(chunks, dim=0)
+    if merged.shape == orig_shape:
+        return merged
+    return merged[: orig_shape[0]]

coreml_suite/lcm/__init__.py

@@ -1,8 +1,3 @@
-"""LCM runtime support (sampler-side).
+from .nodes import 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``.
-"""
+__all__ = ["COREML_CONVERT_LCM"]

coreml_suite/lcm/converter.py

@@ -0,0 +1,297 @@
+import os
+import shutil
+import logging
+import time
+import gc
+
+import numpy as np
+import torch
+from diffusers import UNet2DConditionModel, LCMScheduler
+from diffusers.loaders import LoraLoaderMixin
+
+from comfy.model_management import get_torch_device
+from coreml_suite.lcm.unet import UNet2DConditionModelLCM
+
+from transformers import CLIPTextModel
+import coremltools as ct
+
+from folder_paths import get_folder_paths
+
+logging.basicConfig()
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.DEBUG)
+
+MODEL_VERSION = "SimianLuo/LCM_Dreamshaper_v7"
+MODEL_NAME = MODEL_VERSION.split("/")[-1] + "_4k"
+
+import python_coreml_stable_diffusion.unet as unet
+
+unet.ATTENTION_IMPLEMENTATION_IN_EFFECT = unet.AttentionImplementations.SPLIT_EINSUM
+
+
+def get_unets():
+    ref_unet = UNet2DConditionModel.from_pretrained(
+        MODEL_VERSION,
+        subfolder="unet",
+        device_map=None,
+        low_cpu_mem_usage=False,
+    )
+
+    cml_unet = UNet2DConditionModelLCM.from_config(ref_unet.config).eval()
+    cml_unet.load_state_dict(ref_unet.state_dict(), strict=False)
+
+    return cml_unet, ref_unet
+
+
+def get_encoder_hidden_states_shape(unet_config, batch_size):
+    text_encoder = CLIPTextModel.from_pretrained(
+        MODEL_VERSION, subfolder="text_encoder"
+    )
+
+    text_token_sequence_length = text_encoder.config.max_position_embeddings
+    hidden_size = (text_encoder.config.hidden_size,)
+
+    encoder_hidden_states_shape = (
+        batch_size,
+        unet_config.cross_attention_dim or hidden_size,
+        1,
+        text_token_sequence_length,
+    )
+
+    return encoder_hidden_states_shape
+
+
+def get_scheduler():
+    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):
+    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 compile_coreml_model(source_model_path, output_dir, final_name):
+    """Compiles Core ML models using the coremlcompiler utility from Xcode toolchain"""
+    target_path = os.path.join(output_dir, f"{final_name}.mlmodelc")
+    if os.path.exists(target_path):
+        logger.warning(f"Found existing compiled model at {target_path}! Skipping..")
+        return target_path
+
+    logger.info(f"Compiling {source_model_path}")
+    source_model_name = os.path.basename(os.path.splitext(source_model_path)[0])
+
+    os.system(f"xcrun coremlcompiler compile {source_model_path} {output_dir}")
+    compiled_output = os.path.join(output_dir, f"{source_model_name}.mlmodelc")
+    shutil.move(compiled_output, target_path)
+
+    return target_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):
+    from python_coreml_stable_diffusion.unet import calculate_conv2d_output_shape
+
+    additional_residuals_shapes = []
+
+    batch_size = sample_shape[0]
+    h, w = sample_shape[2:]
+
+    # conv_in
+    out_h, out_w = calculate_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 = calculate_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}")
+
+
+def compile_model(out_path, out_name):
+    # Compile the model
+    target_path = compile_coreml_model(
+        out_path, get_folder_paths("unet")[0], f"{out_name}_unet"
+    )
+    logger.info(f"Compiled {out_path} to {target_path}")
+    return target_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)
+    compile_model(out_path=out_path, out_name=out_name)

coreml_suite/lcm/nodes.py

@@ -0,0 +1,70 @@
+import os
+
+from coremltools import ComputeUnit
+from python_coreml_stable_diffusion.coreml_model import CoreMLModel
+
+from coreml_suite import COREML_NODE
+from coreml_suite.lcm import converter as lcm_converter
+
+
+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.
+        """
+        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,
+            )
+        target_path = lcm_converter.compile_model(out_path=out_path, out_name=out_name)
+
+        return (CoreMLModel(target_path, compute_unit, "compiled"),)

coreml_suite/lcm/unet.py

@@ -0,0 +1,99 @@
+from overrides import overrides
+from python_coreml_stable_diffusion.unet import UNet2DConditionModel, 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
+
+    @overrides(check_signature=False)
+    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,)

coreml_suite/models.py

@@ -1,44 +1,15 @@
-"""Framework-coupled glue between Core ML UNets and ComfyUI's sampler stack.
-
-Pure math (CoreMLInputs, SDXL detection, time_ids/text_embeds assembly,
-sdxl_model_function_wrapper) lives in coreml_suite.core.*.
-This module is what touches comfy.*: model_base, ModelPatcher, the
-diffusion_model wrapper, and the maintainer-facing add_sdxl_model_options
-adapter.
-"""
+import numpy as np
 import torch
 
 from comfy import model_base
 from comfy.model_management import get_torch_device
 from comfy.model_patcher import ModelPatcher
-
 from coreml_suite.config import get_model_config, ModelVersion
-from coreml_suite.core.inputs import CoreMLInputs
-from coreml_suite.core.latents import merge_chunks
-from coreml_suite.core.sdxl import (
-    build_sdxl_text_embeds,
-    build_sdxl_time_ids,
-    is_sdxl,
-    is_sdxl_base,
-    is_sdxl_refiner,
-    sdxl_model_function_wrapper,
-)
+from coreml_suite.controlnet import extract_residual_kwargs, chunk_control
+from coreml_suite.latents import chunk_batch, merge_chunks
 from coreml_suite.lcm.utils import is_lcm
 from coreml_suite.logger import logger
 
-__all__ = [
-    "CoreMLInputs",
-    "CoreMLModelWrapper",
-    "CoreMLModelWrapperLCM",
-    "add_sdxl_model_options",
-    "get_latent_image",
-    "get_model_patcher",
-    "is_sdxl",
-    "is_sdxl_base",
-    "is_sdxl_refiner",
-    "sdxl_model_function_wrapper",
-]
-
 
 class CoreMLModelWrapper:
     def __init__(self, coreml_model):
@@ -97,27 +68,204 @@ class CoreMLModelWrapperLCM(CoreMLModelWrapper):
         self.config = None
 
 
+class CoreMLInputs:
+    def __init__(self, x, t, context, control, **kwargs):
+        self.x = x
+        self.t = t
+        self.context = context
+        self.control = control
+        self.time_ids = kwargs.get("time_ids")
+        self.text_embeds = kwargs.get("text_embeds")
+        self.ts_cond = kwargs.get("timestep_cond")
+
+    def coreml_kwargs(self, expected_inputs):
+        sample = self.x.cpu().numpy().astype(np.float16)
+
+        context = self.context.cpu().numpy().astype(np.float16)
+        context = context.transpose(0, 2, 1)[:, :, None, :]
+
+        t = self.t.cpu().numpy().astype(np.float16)
+
+        model_input_kwargs = {
+            "sample": sample,
+            "encoder_hidden_states": context,
+            "timestep": t,
+        }
+        residual_kwargs = extract_residual_kwargs(expected_inputs, self.control)
+        model_input_kwargs |= residual_kwargs
+
+        # LCM
+        if self.ts_cond is not None:
+            model_input_kwargs["timestep_cond"] = (
+                self.ts_cond.cpu().numpy().astype(np.float16)
+            )
+
+        # SDXL
+        if "text_embeds" in expected_inputs:
+            model_input_kwargs["text_embeds"] = (
+                self.text_embeds.cpu().numpy().astype(np.float16)
+            )
+        if "time_ids" in expected_inputs:
+            model_input_kwargs["time_ids"] = (
+                self.time_ids.cpu().numpy().astype(np.float16)
+            )
+
+        return model_input_kwargs
+
+    def chunks(self, expected_inputs):
+        sample_shape = expected_inputs["sample"]["shape"]
+        timestep_shape = expected_inputs["timestep"]["shape"]
+        hidden_shape = expected_inputs["encoder_hidden_states"]["shape"]
+        context_shape = (hidden_shape[0], hidden_shape[3], hidden_shape[1])
+
+        chunked_x = chunk_batch(self.x, sample_shape)
+        ts = list(torch.full((len(chunked_x), timestep_shape[0]), self.t[0]))
+        chunked_context = chunk_batch(self.context, context_shape)
+
+        chunked_control = [None] * len(chunked_x)
+        if self.control is not None:
+            chunked_control = chunk_control(self.control, sample_shape[0])
+
+        chunked_ts_cond = [None] * len(chunked_x)
+        if self.ts_cond is not None:
+            ts_cond_shape = expected_inputs["timestep_cond"]["shape"]
+            chunked_ts_cond = chunk_batch(self.ts_cond, ts_cond_shape)
+
+        chunked_time_ids = [None] * len(chunked_x)
+        if expected_inputs.get("time_ids") is not None:
+            time_ids_shape = expected_inputs["time_ids"]["shape"]
+            if self.time_ids is None:
+                self.time_ids = torch.zeros(len(chunked_x), *time_ids_shape[1:]).to(
+                    self.x.device
+                )
+            chunked_time_ids = chunk_batch(self.time_ids, time_ids_shape)
+
+        chunked_text_embeds = [None] * len(chunked_x)
+        if expected_inputs.get("text_embeds") is not None:
+            text_embeds_shape = expected_inputs["text_embeds"]["shape"]
+            if self.text_embeds is None:
+                self.text_embeds = torch.zeros(
+                    len(chunked_x), *text_embeds_shape[1:]
+                ).to(self.x.device)
+            chunked_text_embeds = chunk_batch(self.text_embeds, text_embeds_shape)
+
+        return [
+            CoreMLInputs(
+                x,
+                t,
+                context,
+                control,
+                timestep_cond=ts_cond,
+                time_ids=time_ids,
+                text_embeds=text_embeds,
+            )
+            for x, t, context, control, ts_cond, time_ids, text_embeds in zip(
+                chunked_x,
+                ts,
+                chunked_context,
+                chunked_control,
+                chunked_ts_cond,
+                chunked_time_ids,
+                chunked_text_embeds,
+            )
+        ]
+
+
+def is_sdxl(coreml_model):
+    return (
+        "time_ids" in coreml_model.expected_inputs
+        and "text_embeds" in coreml_model.expected_inputs
+    )
+
+
+def is_sdxl_base(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 6
+    )
+
+
+def is_sdxl_refiner(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 5
+    )
+
+
+def sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False):
+    def wrapper(model_function, params):
+        x = params["input"]
+        t = params["timestep"]
+        c = params["c"]
+
+        context = c.get("c_crossattn")
+
+        if context is None:
+            return torch.zeros_like(x)
+
+        if refiner and context is not None:
+            # converted refiner accepts only g clip
+            c["c_crossattn"] = context[:, :, 768:]
+
+        return model_function(x, t, **c, time_ids=time_ids, text_embeds=text_embeds)
+
+    return wrapper
+
+
 def add_sdxl_model_options(model_patcher, positive, negative):
     mp = model_patcher.clone()
 
     pos_dict = positive[0][1]
     neg_dict = negative[0][1]
 
-    is_base = model_patcher.model.diffusion_model.is_sdxl_base
+    pos_pooled = pos_dict["pooled_output"]
+    neg_pooled = neg_dict["pooled_output"]
+
+    pos_time_ids = [
+        pos_dict.get("height", 768),
+        pos_dict.get("width", 768),
+        pos_dict.get("crop_h", 0),
+        pos_dict.get("crop_w", 0),
+    ]
+
+    neg_time_ids = [
+        neg_dict.get("height", 768),
+        neg_dict.get("width", 768),
+        neg_dict.get("crop_h", 0),
+        neg_dict.get("crop_w", 0),
+    ]
+
+    if model_patcher.model.diffusion_model.is_sdxl_base:
+        pos_time_ids += [
+            pos_dict.get("target_height", 768),
+            pos_dict.get("target_width", 768),
+        ]
+
+        neg_time_ids += [
+            neg_dict.get("target_height", 768),
+            neg_dict.get("target_width", 768),
+        ]
+
     is_refiner = model_patcher.model.diffusion_model.is_sdxl_refiner
+    if is_refiner:
+        pos_time_ids += [
+            pos_dict.get("aesthetic_score", 6),
+        ]
 
-    time_ids = build_sdxl_time_ids(
-        pos_dict, neg_dict, is_base=is_base, is_refiner=is_refiner
-    )
-    text_embeds = build_sdxl_text_embeds(
-        pos_dict["pooled_output"], neg_dict["pooled_output"]
-    )
+        neg_time_ids += [
+            neg_dict.get("aesthetic_score", 2.5),
+        ]
 
-    mp.model_options |= {
+    time_ids = torch.tensor([pos_time_ids, neg_time_ids])
+    text_embeds = torch.cat((pos_pooled, neg_pooled))
+
+    model_options = {
         "model_function_wrapper": sdxl_model_function_wrapper(
             time_ids, text_embeds, is_refiner
         ),
     }
+    mp.model_options |= model_options
+
     return mp
 
 

coreml_suite/nodes.py

@@ -1,10 +1,13 @@
 import os
 
 from coremltools import ComputeUnit
+from python_coreml_stable_diffusion.coreml_model import CoreMLModel
+from python_coreml_stable_diffusion.unet import AttentionImplementations
 
 import folder_paths
 from coreml_suite import COREML_NODE
-from coreml_suite.coreml_model import CoreMLModel
+from coreml_suite import converter
+from coreml_suite.config import ModelVersion
 from coreml_suite.lcm.utils import add_lcm_model_options, lcm_patch, is_lcm
 from coreml_suite.logger import logger
 from nodes import KSampler, LoraLoader, KSamplerAdvanced
@@ -17,26 +20,6 @@ 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):
@@ -180,7 +163,7 @@ class CoreMLLoader(COREML_NODE):
 
     @classmethod
     def coreml_filenames(cls):
-        extensions = (".mlpackage",)
+        extensions = (".mlmodelc", ".mlpackage")
         all_paths = folder_paths.get_filename_list_(cls.PACKAGE_DIRNAME)[1]
         coreml_paths = folder_paths.filter_files_extensions(all_paths, extensions)
 
@@ -191,7 +174,9 @@ class CoreMLLoader(COREML_NODE):
 
         coreml_path = self.coreml_filenames()[coreml_name]
 
-        return (CoreMLModel(coreml_path, compute_unit),)
+        sources = "compiled" if coreml_name.endswith(".mlmodelc") else "packages"
+
+        return (CoreMLModel(coreml_path, compute_unit, sources),)
 
 
 class CoreMLLoaderUNet(CoreMLLoader):
@@ -225,26 +210,28 @@ class CoreMLModelAdapter(COREML_NODE):
 
 
 class CoreMLConverter(COREML_NODE):
-    """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.
-    """
+    """Converts a LCM model to Core ML."""
 
     @classmethod
     def INPUT_TYPES(cls):
         return {
             "required": {
                 "ckpt_name": (folder_paths.get_filename_list("checkpoints"),),
-                "height": ("INT", {"default": 512, "min": 8, "step": 8}),
-                "width": ("INT", {"default": 512, "min": 8, "step": 8}),
+                "model_version": (
+                    [
+                        ModelVersion.SD15.name,
+                        ModelVersion.SDXL.name,
+                    ],
+                ),
+                "height": ("INT", {"default": 512, "min": 256, "max": 2048, "step": 8}),
+                "width": ("INT", {"default": 512, "min": 256, "max": 2048, "step": 8}),
                 "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                 "attention_implementation": (
-                    _discover(
-                        "list_attention_impls",
-                        ["SPLIT_EINSUM", "SPLIT_EINSUM_V2", "ORIGINAL"],
-                    ),
+                    [
+                        AttentionImplementations.SPLIT_EINSUM.name,
+                        AttentionImplementations.SPLIT_EINSUM_V2.name,
+                        AttentionImplementations.ORIGINAL.name,
+                    ],
                 ),
                 "compute_unit": (
                     [
@@ -257,15 +244,6 @@ class CoreMLConverter(COREML_NODE):
                 "controlnet_support": ("BOOLEAN", {"default": False}),
             },
             "optional": {
-                # k-means weight palettization. Kept optional so workflows
-                # that omit it still validate — ComfyUI rejects a prompt that
-                # omits any `required` input. When omitted it defaults to
-                # "none", identical to unquantized behavior and filename, so
-                # existing cached .mlpackages still resolve.
-                "quantize_nbits": (
-                    _discover("list_quant_modes", ["none", "8", "6", "4"]),
-                    {"default": "none"},
-                ),
                 "lora_params": ("LORA_PARAMS",),
             },
         }
@@ -277,20 +255,18 @@ class CoreMLConverter(COREML_NODE):
     def convert(
         self,
         ckpt_name,
+        model_version,
         height,
         width,
         batch_size,
         attention_implementation,
         compute_unit,
         controlnet_support,
-        quantize_nbits="none",
         lora_params=None,
     ):
-        """Converts a checkpoint's UNet to Core ML.
+        """Converts a LCM model 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.
@@ -300,8 +276,10 @@ 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 "Load Core ML UNet" node.
+        can be loaded with the "LCMCoreMLLoaderUNet" 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])
@@ -310,19 +288,24 @@ class CoreMLConverter(COREML_NODE):
         h = height
         w = width
         sample_size = (h // 8, w // 8)
-        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=coreml_diffusion.lora_names_from_params(lora_params),
-            quantize_nbits=quantize_nbits,
+        batch_size = batch_size
+        cn_support_str = "_cn" if controlnet_support else ""
+        lora_str = (
+            "_" + "_".join(lora_param[0].split(".")[0] for lora_param in lora_params)
+            if lora_params
+            else ""
         )
 
+        attn_str = (
+            "_"
+            + {"SPLIT_EINSUM": "se", "SPLIT_EINSUM_V2": "se2", "ORIGINAL": "orig"}[
+                attention_implementation
+            ]
+        )
+
+        out_name = f"{ckpt_name.split('.')[0]}{lora_str}_{batch_size}x{w}x{h}{cn_support_str}{attn_str}"
+        out_name = out_name.replace(" ", "_")
+
         logger.info(f"Converting {ckpt_name} to {out_name}")
         logger.info(f"Batch size: {batch_size}")
         logger.info(f"Width: {w}, Height: {h}")
@@ -330,14 +313,11 @@ class CoreMLConverter(COREML_NODE):
         logger.info(f"Attention implementation: {attention_implementation}")
 
         if lora_params:
-            logger.info("LoRAs used:")
+            logger.info(f"LoRAs used:")
             for lora_param in lora_params:
                 logger.info(f"  {lora_param[0]} - strength: {lora_param[1]}")
 
-        # Resolve the ComfyUI models/unet path here (a node concern); the package
-        # takes the output path as an injected argument.
-        unet_path = folder_paths.get_folder_paths("unet")[0]
-        unet_out_path = os.path.join(unet_path, f"{out_name}_unet.mlpackage")
+        unet_out_path = converter.get_out_path("unet", f"{out_name}")
         ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
 
         config_filename = ckpt_name.split(".")[0] + ".yaml"
@@ -345,19 +325,22 @@ class CoreMLConverter(COREML_NODE):
         if config_path:
             logger.info(f"Using config file {config_path}")
 
-        coreml_diffusion.convert(
-            ckpt_path,
-            None,  # model_version auto-detected from the checkpoint
-            unet_out_path,
+        converter.convert(
+            ckpt_path=ckpt_path,
+            model_version=model_version,
+            unet_out_path=unet_out_path,
             sample_size=sample_size,
             batch_size=batch_size,
             controlnet_support=controlnet_support,
             lora_weights=lora_weights,
             attn_impl=attention_implementation,
             config_path=config_path,
-            quantize_nbits=quantize_nbits,
         )
-        return (CoreMLModel(unet_out_path, compute_unit),)
+        unet_target_path = converter.compile_model(
+            out_path=unet_out_path, out_name=out_name, submodule_name="unet"
+        )
+
+        return (CoreMLModel(unet_target_path, compute_unit, "compiled"),)
 
     @staticmethod
     def lora_path(lora_name):

docs/conversion.md

@@ -1,93 +0,0 @@
-# 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.

docs/faq.md

@@ -1,77 +0,0 @@
-# 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).

docs/hardware.md

@@ -1,95 +0,0 @@
-# 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.

docs/limitations.md

@@ -1,53 +0,0 @@
-# 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).

docs/nodes.md

@@ -1,157 +0,0 @@
-# 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`.

docs/troubleshooting.md

@@ -1,86 +0,0 @@
-# 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.

docs/workflows.md

@@ -1,103 +0,0 @@
-# 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)

pyproject.toml

@@ -1,63 +0,0 @@
-[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 = "2.1.2"
-license = "MIT"
-requires-python = ">=3.12"
-dependencies = [
-    # 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",
-]
-
-[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 = "https://raw.githubusercontent.com/aszc-dev/ComfyUI-CoreMLSuite/main/assets/snake.png"
-requires-comfyui = ">=0.3.27"
-
-[dependency-groups]
-dev = [
-    "pillow>=12.2.0",
-    "psutil>=7.2.2",
-    "pytest>=9.0.3",
-]
-comfy = [
-    "comfyui-frontend-package==1.14.6",
-    "torchvision",
-    "torchaudio",
-    "torchsde",
-    "einops",
-    "tokenizers>=0.13.3",
-    "safetensors>=0.4.2",
-    "aiohttp>=3.11.8",
-    "yarl>=1.18.0",
-    "kornia>=0.7.1",
-    "spandrel",
-    "soundfile",
-    "sentencepiece",
-]
-
-[tool.pytest.ini_options]
-markers = [
-    "unit: framework-free unit test (Tier 0)",
-    "smoke: macOS-ARM smoke test on a synthetic micro-model (Tier 1)",
-    "m2: requires Apple Silicon + Neural Engine (Tier 2)",
-]
-testpaths = ["tests"]
-addopts = ["--import-mode=importlib", "--confcutdir=tests"]

requirements.txt

@@ -1,4 +1,6 @@
-coreml-diffusion>=0.1.4,<0.2
-coremltools>=9,<10
-numpy>=2,<3
-diffusers>=0.30
+git+https://github.com/apple/ml-stable-diffusion.git
+coremltools>=7.1
+overrides
+diffusers>=0.22
+peft>=0.6.2
+omegaconf>=2.3

tests/conftest.py

@@ -1,61 +0,0 @@
-"""Pytest bootstrap for ComfyUI-CoreMLSuite tests.
-
-- Adds the ComfyUI checkout to sys.path so the framework-coupled modules
-  that transitively import `comfy.*` resolve when pytest is invoked from
-  this package's root.
-- Auto-applies tier markers based on the directory a test lives in, so
-  individual files don't have to repeat @pytest.mark.unit / .smoke.
-"""
-import sys
-from pathlib import Path
-
-import pytest
-
-REPO_ROOT = Path(__file__).resolve().parents[1]
-COMFY_DIR = REPO_ROOT.parents[1]
-
-for p in (str(COMFY_DIR), str(REPO_ROOT)):
-    if p not in sys.path:
-        sys.path.insert(0, p)
-
-
-_TIER_BY_DIR = {
-    "tests/unit": "unit",
-    "tests/m2": "m2",
-    "tests/integration": "m2",
-    "tests/smoke": "smoke",
-}
-
-# When the user asks for a single tier (-m unit / -m smoke), skip the other
-# directories at collection time. Tier-0 cannot afford to import tests/smoke
-# files because they pull in coremltools which Linux CI won't have.
-_TIER_DIRS = {
-    "unit": ("/tests/unit/",),
-    "m2": ("/tests/m2/", "/tests/integration/"),
-    "smoke": ("/tests/smoke/",),
-}
-
-
-def pytest_ignore_collect(collection_path, config):
-    expr = config.option.markexpr
-    if expr not in _TIER_DIRS:
-        return None
-    allowed = _TIER_DIRS[expr]
-    rel = str(collection_path).replace("\\", "/")
-    if "/tests/" not in rel:
-        return None
-    # Always allow tests/ root + the tier's own dirs.
-    if rel.endswith("/tests"):
-        return None
-    if any(frag in rel + "/" for frag in allowed):
-        return None
-    return True
-
-
-def pytest_collection_modifyitems(config, items):
-    for item in items:
-        path = str(item.fspath).replace("\\", "/")
-        for fragment, marker in _TIER_BY_DIR.items():
-            if f"/{fragment}/" in path:
-                item.add_marker(getattr(pytest.mark, marker))
-                break

tests/integration/test_basic_conversion_1_5.py

@@ -0,0 +1,72 @@
+import json
+import os
+
+import pytest
+import requests
+from PIL import Image
+import numpy as np
+
+from folder_paths import get_save_image_path, get_output_directory
+
+IMAGE_PREFIX = "E2E-1.5-CoreML"
+
+
+class OutputImageRepository:
+    def __init__(self, name_prefix):
+        self.name_prefix = name_prefix
+
+    def list_images(self):
+        full_output_folder, _, _, _, _ = get_save_image_path(
+            self.name_prefix, get_output_directory(), 512, 512
+        )
+        return full_output_folder, os.listdir(full_output_folder)
+
+    def delete_images(self):
+        full_output_folder, images = self.list_images()
+        for image in images:
+            os.remove(os.path.join(full_output_folder, image))
+
+
+@pytest.fixture(scope="module")
+def output_image_repository():
+    repo = OutputImageRepository(IMAGE_PREFIX)
+    yield repo
+    repo.delete_images()
+
+
+def test_basic_conversion_1_5(output_image_repository):
+    with open("integration/workflows/e2e-1.5-basic-conversion.json") as f:
+        prompt = json.load(f)
+    queue_prompt(prompt)
+
+    full_output_folder, images = output_image_repository.list_images()
+    assert len(images) == 2
+    assert all(image.startswith(IMAGE_PREFIX) for image in images)
+    assert all(image.endswith(".png") for image in images)
+    assert all(
+        os.path.isfile(os.path.join(full_output_folder, image)) for image in images
+    )
+
+    image1 = Image.open(os.path.join(full_output_folder, images[0]))
+    image2 = Image.open(os.path.join(full_output_folder, images[1]))
+    assert psnr(np.array(image1), np.array(image2)) > 30
+    assert psnr(np.array(image2), np.array(image1)) > 30
+
+
+def psnr(img1, img2):
+    mse = np.mean((img1 - img2) ** 2)
+    if mse == 0:
+        return 100
+    PIXEL_MAX = 255.0
+    return 20 * np.log10(PIXEL_MAX / np.sqrt(mse))
+
+
+def queue_prompt(prompt: dict):
+    p = {"prompt": prompt}
+    data = json.dumps(p).encode("utf-8")
+    req = requests.post("http://localhost:8188/prompt", data=data)
+    assert req.status_code == 200
+    while True:
+        req = requests.get("http://localhost:8188/prompt")
+        if req.json()["exec_info"]["queue_remaining"] == 0:
+            break

tests/integration/workflows/e2e-1.5-basic-conversion.json

@@ -107,6 +107,7 @@
   "10": {
     "inputs": {
       "ckpt_name": "dreamshaper_8.safetensors",
+      "model_version": "SD15",
       "height": 512,
       "width": 512,
       "batch_size": 1,
@@ -178,4 +179,4 @@
       "title": "Save Image"
     }
   }
-}
+}

tests/m2/goldens/sd15_seed42.sha256

@@ -1 +0,0 @@
-e89344e544d4edfbd3ebe9a1c78dadb2729f53549666052b74ac7308f326f4fc

tests/m2/test_golden_image.py

@@ -1,170 +0,0 @@
-"""[M2-ANE] golden-image anchor.
-
-Runs the e2e SD1.5 + CoreML workflow against a local ComfyUI server, fetches
-the generated PNG, and asserts both:
-  - byte-identical SHA256 against the stored golden, OR
-  - PSNR >= GOLDEN_PSNR_MIN_DB against the stored golden PNG.
-
-The hash is the strict gate (a refactor that doesn't touch the math
-should hit it). PSNR is the soft gate that tolerates the drift a
-toolchain bump injects through different MIL graphs / kernel selection
-/ fp accumulation order — anything below the threshold is treated as a
-regression.
-
-The 20 dB default absorbs Apple Neural Engine run-to-run nondeterminism:
-the same model and seed can drift several dB between runs as the 20
-sampling steps amplify tiny per-step UNet differences (kernel selection /
-fp accumulation order). Same-scene ANE outputs have been observed at
-~23 dB, so 20 leaves margin while still catching gross regressions — a
-broken image lands far lower. Bump it up for pure-refactor PRs that must
-not change math; down for toolchain bumps.
-
-Skips entirely on non-Apple-Silicon hosts or when the server / converted
-model is missing, so the unit lane on Linux still passes.
-
-The first run with no golden writes one and fails so it's reviewed before
-being committed.
-"""
-import hashlib
-import json
-import os
-import platform
-import shutil
-import time
-import urllib.error
-import urllib.request
-from pathlib import Path
-
-import numpy as np
-import pytest
-from PIL import Image
-
-REPO_ROOT = Path(__file__).resolve().parents[2]
-COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
-COMFY_HOST = os.environ.get("COMFY_HOST", "localhost")
-COMFY_PORT = int(os.environ.get("COMFY_PORT", "8188"))
-COMFY_URL = f"http://{COMFY_HOST}:{COMFY_PORT}"
-
-CKPT_NAME = os.environ.get("CKPT_NAME", "v1-5-pruned-emaonly.safetensors")
-WORKFLOW_PATH = (
-    REPO_ROOT / "tests" / "integration" / "workflows" / "e2e-1.5-basic-conversion.json"
-)
-GOLDEN_DIR = Path(__file__).parent / "goldens"
-GOLDEN_HASH_PATH = GOLDEN_DIR / "sd15_seed42.sha256"
-GOLDEN_PNG_PATH = GOLDEN_DIR / "sd15_seed42.png"
-GOLDEN_PSNR_MIN_DB = float(os.environ.get("GOLDEN_PSNR_MIN_DB", "20"))
-SEED = 42
-
-
-def _server_reachable() -> bool:
-    try:
-        with urllib.request.urlopen(f"{COMFY_URL}/prompt", timeout=3) as r:
-            return r.status == 200
-    except (urllib.error.URLError, urllib.error.HTTPError, ConnectionError):
-        return False
-
-
-@pytest.fixture(scope="module")
-def comfy_server():
-    if platform.machine() != "arm64":
-        pytest.skip("requires Apple Silicon")
-    if not _server_reachable():
-        pytest.skip(f"ComfyUI server not reachable at {COMFY_URL}")
-    return COMFY_URL
-
-
-def _http_post_json(path: str, payload: dict) -> dict:
-    data = json.dumps(payload).encode("utf-8")
-    req = urllib.request.Request(
-        f"{COMFY_URL}{path}", data=data,
-        headers={"Content-Type": "application/json"}, method="POST",
-    )
-    with urllib.request.urlopen(req, timeout=300) as r:
-        return json.loads(r.read().decode())
-
-
-def _http_get_json(path: str, timeout: int = 300) -> dict:
-    """ComfyUI runs UNet inference on its single asyncio loop, so GET /prompt
-    blocks while the queued prompt is executing. Use a generous timeout."""
-    with urllib.request.urlopen(f"{COMFY_URL}{path}", timeout=timeout) as r:
-        return json.loads(r.read().decode())
-
-
-def _drain_queue(timeout_s: int = 600) -> None:
-    deadline = time.time() + timeout_s
-    while time.time() < deadline:
-        try:
-            q = _http_get_json("/prompt")
-        except (urllib.error.URLError, TimeoutError):
-            # Transient block while server executes; retry until our overall
-            # deadline expires.
-            continue
-        if q.get("exec_info", {}).get("queue_remaining", -1) == 0:
-            return
-        time.sleep(2)
-    raise TimeoutError(f"queue did not drain within {timeout_s}s")
-
-
-def _post_workflow_and_collect_png() -> bytes:
-    workflow = json.loads(WORKFLOW_PATH.read_text())
-    for nid in ("4", "10"):
-        if nid in workflow:
-            workflow[nid]["inputs"]["ckpt_name"] = CKPT_NAME
-    for nid in ("3", "11"):
-        if nid in workflow and "seed" in workflow[nid].get("inputs", {}):
-            workflow[nid]["inputs"]["seed"] = SEED
-    # Drop the MPS reference branch — only the Core ML pipeline is needed here.
-    for nid in ("3", "8", "9"):
-        workflow.pop(nid, None)
-
-    _http_post_json("/prompt", {"prompt": workflow})
-    _drain_queue()
-
-    comfy_out = COMFY_DIR / "output"
-    matches = sorted(comfy_out.glob("E2E-1.5-CoreML_*.png"), reverse=True)
-    if not matches:
-        raise FileNotFoundError(f"no Core ML image under {comfy_out}")
-    return matches[0].read_bytes()
-
-
-def _psnr(a: np.ndarray, b: np.ndarray) -> float:
-    mse = float(np.mean((a.astype(np.float64) - b.astype(np.float64)) ** 2))
-    if mse == 0:
-        return 100.0
-    return 20.0 * float(np.log10(255.0 / np.sqrt(mse)))
-
-
-def test_sd15_seed42_image_matches_golden(comfy_server):
-    GOLDEN_DIR.mkdir(parents=True, exist_ok=True)
-    png_bytes = _post_workflow_and_collect_png()
-    sha = hashlib.sha256(png_bytes).hexdigest()
-
-    if not GOLDEN_HASH_PATH.exists() or not GOLDEN_PNG_PATH.exists():
-        GOLDEN_HASH_PATH.write_text(sha + "\n")
-        # Persist the PNG too for visual diffing + PSNR.
-        tmp_path = Path(__file__).parent / "_latest_generated.png"
-        tmp_path.write_bytes(png_bytes)
-        shutil.copy2(tmp_path, GOLDEN_PNG_PATH)
-        pytest.fail(
-            f"No golden present; wrote {GOLDEN_HASH_PATH.name} and "
-            f"{GOLDEN_PNG_PATH.name}. Review the image and re-run."
-        )
-
-    expected_hash = GOLDEN_HASH_PATH.read_text().strip()
-    if sha == expected_hash:
-        return
-
-    # Hash drift: fall back to PSNR to distinguish a refactor-safe rounding
-    # change from a real regression.
-    a = np.array(Image.open(GOLDEN_PNG_PATH).convert("RGB"))
-    b_path = Path(__file__).parent / "_latest_generated.png"
-    b_path.write_bytes(png_bytes)
-    b = np.array(Image.open(b_path).convert("RGB"))
-    if a.shape != b.shape:
-        pytest.fail(f"shape mismatch: golden={a.shape} actual={b.shape}")
-    psnr_db = _psnr(a, b)
-    assert psnr_db >= GOLDEN_PSNR_MIN_DB, (
-        f"hash drifted (got {sha[:12]}.., expected {expected_hash[:12]}..) and "
-        f"PSNR {psnr_db:.2f} dB < {GOLDEN_PSNR_MIN_DB} dB threshold; "
-        f"diff PNG at {b_path}"
-    )

tests/unit/test_characterization_controlnet.py

@@ -1,186 +0,0 @@
-"""Characterization tests for coreml_suite.controlnet.
-
-Locks shapes + dtypes + zero-fill behavior of expand_inputs / no_control /
-extract_residual_kwargs / chunk_control. These pure helpers feed the Core ML
-UNet's additional_residual_N inputs; any drift here silently breaks
-ControlNet-based workflows.
-"""
-import numpy as np
-import pytest
-import torch
-
-from coreml_suite.core.controlnet import (
-    chunk_control,
-    expand_inputs,
-    extract_residual_kwargs,
-    no_control,
-)
-
-
-@pytest.fixture(autouse=True)
-def _deterministic_seed():
-    torch.manual_seed(0)
-    np.random.seed(0)
-
-
-SD15_RESIDUAL_SPEC = {
-    "additional_residual_0": {"shape": (2, 320, 64, 64)},
-    "additional_residual_1": {"shape": (2, 640, 32, 32)},
-    "additional_residual_2": {"shape": (2, 1280, 8, 8)},
-}
-NON_RESIDUAL_SPEC = {
-    "sample": {"shape": (2, 4, 64, 64)},
-    "encoder_hidden_states": {"shape": (2, 77, 768)},
-}
-
-
-# ---------- expand_inputs ----------------------------------------------------
-
-
-def test_expand_inputs_doubles_singleton_numpy():
-    inputs = {"a": np.ones((1, 4), dtype=np.float32)}
-    out = expand_inputs(inputs)
-    assert out["a"].shape == (2, 4)
-    assert np.array_equal(out["a"], np.ones((2, 4)))
-
-
-def test_expand_inputs_doubles_singleton_torch():
-    inputs = {"a": torch.ones(1, 4)}
-    out = expand_inputs(inputs)
-    assert out["a"].shape == (2, 4)
-    assert torch.equal(out["a"], torch.ones(2, 4))
-
-
-def test_expand_inputs_doubles_singleton_list():
-    inputs = {"a": [42]}
-    out = expand_inputs(inputs)
-    assert out["a"] == [42, 42]
-
-
-def test_expand_inputs_skips_already_batched():
-    """batch > 1 inputs are returned unchanged (same object identity)."""
-    arr = np.ones((2, 4), dtype=np.float32)
-    tensor = torch.ones(3, 4)
-    lst = [1, 2]
-    out = expand_inputs({"a": arr, "b": tensor, "c": lst})
-    assert out["a"] is arr
-    assert out["b"] is tensor
-    assert out["c"] is lst
-
-
-def test_expand_inputs_preserves_unknown_value_types():
-    # Strings/None pass through untouched — locks current permissive contract.
-    inputs = {"s": "hello", "none": None, "int": 7}
-    out = expand_inputs(inputs)
-    assert out == {"s": "hello", "none": None, "int": 7}
-
-
-# ---------- no_control -------------------------------------------------------
-
-
-def test_no_control_returns_zero_fp16_for_residuals():
-    out = no_control({**SD15_RESIDUAL_SPEC, **NON_RESIDUAL_SPEC})
-    # Only additional_residual_* keys are produced.
-    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
-    for key, spec in SD15_RESIDUAL_SPEC.items():
-        arr = out[key]
-        assert arr.shape == spec["shape"]
-        assert arr.dtype == np.float16
-        assert np.all(arr == 0)
-
-
-def test_no_control_returns_empty_when_no_residuals():
-    out = no_control(NON_RESIDUAL_SPEC)
-    assert out == {}
-
-
-# ---------- extract_residual_kwargs -----------------------------------------
-
-
-def test_extract_residual_kwargs_empty_when_model_has_no_residual_inputs():
-    out = extract_residual_kwargs(NON_RESIDUAL_SPEC, control={"output": [], "middle": []})
-    assert out == {}
-
-
-def test_extract_residual_kwargs_none_control_returns_no_control_shapes():
-    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control=None)
-    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
-    for key, spec in SD15_RESIDUAL_SPEC.items():
-        assert out[key].shape == spec["shape"]
-        assert out[key].dtype == np.float16
-        assert np.all(out[key] == 0)
-
-
-def test_extract_residual_kwargs_flattens_output_then_middle_and_casts_fp16():
-    """output residuals come first (indexed 0..N-1), then middle residuals
-    (indexed N..M-1). Values come out of CPU as fp16 numpy arrays."""
-    control = {
-        "output": [torch.ones(2, 320, 64, 64) * 0.5, torch.ones(2, 640, 32, 32) * 2.0],
-        "middle": [torch.ones(2, 1280, 8, 8) * -1.0],
-    }
-    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control)
-    assert set(out.keys()) == {"additional_residual_0", "additional_residual_1", "additional_residual_2"}
-    assert out["additional_residual_0"].shape == (2, 320, 64, 64)
-    assert out["additional_residual_1"].shape == (2, 640, 32, 32)
-    assert out["additional_residual_2"].shape == (2, 1280, 8, 8)
-    for arr in out.values():
-        assert arr.dtype == np.float16
-    # Locked order: index 0 == first output residual (0.5), index 2 == middle (-1.0).
-    assert np.allclose(out["additional_residual_0"], 0.5)
-    assert np.allclose(out["additional_residual_1"], 2.0)
-    assert np.allclose(out["additional_residual_2"], -1.0)
-
-
-# ---------- chunk_control ----------------------------------------------------
-
-
-def test_chunk_control_none_returns_list_of_nones_with_length_target():
-    """`no_control` path: when there's no control, you get [None] * target_size
-    (NOT [None, None] regardless of target — this is the contract today)."""
-    assert chunk_control(None, 1) == [None]
-    assert chunk_control(None, 2) == [None, None]
-    assert chunk_control(None, 4) == [None, None, None, None]
-
-
-@pytest.mark.parametrize(
-    "batch,target,expected_chunks",
-    [(1, 2, 1), (2, 2, 1), (3, 2, 2), (4, 2, 2), (5, 3, 2), (9, 4, 3)],
-)
-def test_chunk_control_shapes_after_chunking(batch, target, expected_chunks):
-    cn = {
-        "output": [
-            torch.randn(batch, 320, 64, 64),
-            torch.randn(batch, 640, 32, 32),
-        ],
-        "middle": [torch.randn(batch, 1280, 8, 8)],
-    }
-    chunks = chunk_control(cn, target)
-    assert len(chunks) == expected_chunks
-    for c in chunks:
-        assert c["output"][0].shape == (target, 320, 64, 64)
-        assert c["output"][1].shape == (target, 640, 32, 32)
-        assert c["middle"][0].shape == (target, 1280, 8, 8)
-
-
-def test_chunk_control_preserves_keys_order():
-    """Output dicts contain exactly {"output", "middle"} in that order."""
-    cn = {
-        "output": [torch.zeros(2, 4, 4, 4)],
-        "middle": [torch.zeros(2, 4, 4, 4)],
-    }
-    chunks = chunk_control(cn, 2)
-    assert list(chunks[0].keys()) == ["output", "middle"]
-
-
-def test_chunk_control_zero_pads_remainder():
-    """A batch=3, target=2 split puts the third row alongside a zero row."""
-    cn = {
-        "output": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
-        "middle": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
-    }
-    chunks = chunk_control(cn, 2)
-    assert len(chunks) == 2
-    last_out = chunks[-1]["output"][0]
-    # First row is the original third row; second row is padding zeros.
-    assert torch.equal(last_out[0], cn["output"][0][2])
-    assert torch.equal(last_out[1], torch.zeros(1, 2, 2))

tests/unit/test_characterization_inputs.py

@@ -1,228 +0,0 @@
-"""Characterization tests for coreml_suite.models.CoreMLInputs.
-
-Locks the shape transforms applied by chunks() and coreml_kwargs() for the
-four model variants the suite supports: SD1.5, LCM (SD1.5 + timestep_cond),
-SDXL base (time_ids len 6), and SDXL refiner (time_ids len 5).
-
-These contracts feed the Core ML UNet at runtime; if a refactor silently
-re-shapes them, generation breaks.
-"""
-import numpy as np
-import pytest
-import torch
-
-from coreml_suite.core.inputs import CoreMLInputs
-
-
-@pytest.fixture(autouse=True)
-def _deterministic_seed():
-    torch.manual_seed(0)
-    np.random.seed(0)
-
-
-# ---------- expected_inputs fixtures (mirror real model expectations) -------
-
-SD15_EXPECTED = {
-    "sample": {"shape": (2, 4, 64, 64)},
-    "timestep": {"shape": (2,)},
-    "encoder_hidden_states": {"shape": (2, 77, 768)},
-}
-
-SD15_WITH_CN = {
-    **SD15_EXPECTED,
-    "additional_residual_0": {"shape": (2, 320, 64, 64)},
-    "additional_residual_1": {"shape": (2, 640, 32, 32)},
-}
-
-LCM_EXPECTED = {
-    **SD15_EXPECTED,
-    "timestep_cond": {"shape": (2, 256)},
-}
-
-SDXL_BASE_EXPECTED = {
-    "sample": {"shape": (2, 4, 128, 128)},
-    "timestep": {"shape": (2,)},
-    "encoder_hidden_states": {"shape": (2, 77, 2048)},
-    "time_ids": {"shape": (2, 6)},
-    "text_embeds": {"shape": (2, 1280)},
-}
-
-SDXL_REFINER_EXPECTED = {
-    "sample": {"shape": (2, 4, 128, 128)},
-    "timestep": {"shape": (2,)},
-    "encoder_hidden_states": {"shape": (2, 77, 1280)},
-    "time_ids": {"shape": (2, 5)},
-    "text_embeds": {"shape": (2, 1280)},
-}
-
-
-def _sd15_inputs(batch=1, with_control=False, with_ts_cond=False):
-    x = torch.randn(batch, 4, 64, 64)
-    t = torch.full((batch,), 999.0)
-    context = torch.randn(batch, 77, 768)
-    control = None
-    if with_control:
-        control = {
-            "output": [torch.randn(batch, 320, 64, 64), torch.randn(batch, 640, 32, 32)],
-            "middle": [],
-        }
-    kwargs = {}
-    if with_ts_cond:
-        kwargs["timestep_cond"] = torch.randn(batch, 256)
-    return CoreMLInputs(x, t, context, control, **kwargs)
-
-
-def _sdxl_inputs(batch=1, refiner=False):
-    x = torch.randn(batch, 4, 128, 128)
-    t = torch.full((batch,), 999.0)
-    ctx_dim = 1280 if refiner else 2048
-    context = torch.randn(batch, 77, ctx_dim)
-    time_ids_dim = 5 if refiner else 6
-    time_ids = torch.randn(batch, time_ids_dim)
-    text_embeds = torch.randn(batch, 1280)
-    return CoreMLInputs(
-        x, t, context, control=None, time_ids=time_ids, text_embeds=text_embeds
-    )
-
-
-# ---------- coreml_kwargs ---------------------------------------------------
-
-
-def test_coreml_kwargs_sd15_shapes_and_fp16():
-    out = _sd15_inputs(batch=1).coreml_kwargs(SD15_EXPECTED)
-    assert set(out.keys()) == {"sample", "encoder_hidden_states", "timestep"}
-    assert out["sample"].shape == (1, 4, 64, 64)
-    assert out["sample"].dtype == np.float16
-    # encoder_hidden_states keeps Comfy's native (b, seq, dim) layout.
-    assert out["encoder_hidden_states"].shape == (1, 77, 768)
-    assert out["encoder_hidden_states"].dtype == np.float16
-    assert out["timestep"].shape == (1,)
-    assert out["timestep"].dtype == np.float16
-
-
-def test_coreml_kwargs_sd15_with_controlnet_emits_residuals():
-    inputs = _sd15_inputs(batch=1, with_control=True)
-    out = inputs.coreml_kwargs(SD15_WITH_CN)
-    assert "additional_residual_0" in out
-    assert "additional_residual_1" in out
-    assert out["additional_residual_0"].shape == (1, 320, 64, 64)
-    assert out["additional_residual_1"].shape == (1, 640, 32, 32)
-
-
-def test_coreml_kwargs_sd15_without_controlnet_zero_fills_residuals():
-    inputs = _sd15_inputs(batch=1, with_control=False)
-    out = inputs.coreml_kwargs(SD15_WITH_CN)
-    assert np.all(out["additional_residual_0"] == 0)
-    assert np.all(out["additional_residual_1"] == 0)
-
-
-def test_coreml_kwargs_lcm_adds_timestep_cond():
-    inputs = _sd15_inputs(batch=1, with_ts_cond=True)
-    out = inputs.coreml_kwargs(LCM_EXPECTED)
-    assert "timestep_cond" in out
-    assert out["timestep_cond"].shape == (1, 256)
-    assert out["timestep_cond"].dtype == np.float16
-
-
-def test_coreml_kwargs_lcm_skips_timestep_cond_when_not_provided():
-    """timestep_cond is only forwarded when the input supplied one — even if
-    the model's expected_inputs lists it."""
-    inputs = _sd15_inputs(batch=1, with_ts_cond=False)
-    out = inputs.coreml_kwargs(LCM_EXPECTED)
-    assert "timestep_cond" not in out
-
-
-def test_coreml_kwargs_sdxl_base_emits_time_ids_and_text_embeds():
-    out = _sdxl_inputs(batch=1, refiner=False).coreml_kwargs(SDXL_BASE_EXPECTED)
-    assert out["time_ids"].shape == (1, 6)
-    assert out["text_embeds"].shape == (1, 1280)
-    assert out["time_ids"].dtype == np.float16
-    assert out["text_embeds"].dtype == np.float16
-
-
-def test_coreml_kwargs_sdxl_refiner_uses_len5_time_ids():
-    out = _sdxl_inputs(batch=1, refiner=True).coreml_kwargs(SDXL_REFINER_EXPECTED)
-    assert out["time_ids"].shape == (1, 5)
-
-
-# ---------- chunks ----------------------------------------------------------
-
-
-def test_chunks_sd15_pad_to_batch2_returns_one_chunk():
-    chunked = _sd15_inputs(batch=1).chunks(SD15_EXPECTED)
-    assert len(chunked) == 1
-    c = chunked[0]
-    assert c.x.shape == (2, 4, 64, 64)
-    assert c.t.shape == (2,)
-    # context shape: (b, seq, dim) padded along batch dim.
-    assert c.context.shape == (2, 77, 768)
-    assert c.control is None
-    assert c.ts_cond is None
-    assert c.time_ids is None
-    assert c.text_embeds is None
-
-
-def test_chunks_sd15_with_controlnet_chunks_residuals_too():
-    chunked = _sd15_inputs(batch=1, with_control=True).chunks(SD15_EXPECTED)
-    assert len(chunked) == 1
-    cn = chunked[0].control
-    assert cn is not None
-    assert cn["output"][0].shape == (2, 320, 64, 64)
-    assert cn["output"][1].shape == (2, 640, 32, 32)
-
-
-def test_chunks_lcm_carries_timestep_cond_per_chunk():
-    chunked = _sd15_inputs(batch=1, with_ts_cond=True).chunks(LCM_EXPECTED)
-    assert len(chunked) == 1
-    assert chunked[0].ts_cond is not None
-    assert chunked[0].ts_cond.shape == (2, 256)
-
-
-def test_chunks_sdxl_base_propagates_time_ids_and_text_embeds():
-    chunked = _sdxl_inputs(batch=1, refiner=False).chunks(SDXL_BASE_EXPECTED)
-    assert len(chunked) == 1
-    c = chunked[0]
-    assert c.time_ids is not None and c.time_ids.shape == (2, 6)
-    assert c.text_embeds is not None and c.text_embeds.shape == (2, 1280)
-
-
-def test_chunks_sdxl_refiner_uses_len5_time_ids():
-    chunked = _sdxl_inputs(batch=1, refiner=True).chunks(SDXL_REFINER_EXPECTED)
-    assert chunked[0].time_ids.shape == (2, 5)
-
-
-def test_chunks_sdxl_synthesizes_zero_time_ids_when_caller_omits():
-    """If the model expects time_ids but caller passed nothing, the suite
-    fabricates a zero-filled tensor. Lock that fallback."""
-    x = torch.randn(1, 4, 128, 128)
-    t = torch.full((1,), 999.0)
-    context = torch.randn(1, 77, 2048)
-    inputs = CoreMLInputs(x, t, context, control=None)
-    chunked = inputs.chunks(SDXL_BASE_EXPECTED)
-    assert chunked[0].time_ids.shape == (2, 6)
-    assert torch.equal(chunked[0].time_ids, torch.zeros(2, 6))
-    assert chunked[0].text_embeds.shape == (2, 1280)
-    assert torch.equal(chunked[0].text_embeds, torch.zeros(2, 1280))
-
-
-def test_chunks_splits_batch_into_multiple_target2_chunks():
-    """batch=5 with target_batch=2 -> 3 chunks (last padded)."""
-    chunked = _sd15_inputs(batch=5).chunks(SD15_EXPECTED)
-    assert len(chunked) == 3
-    for c in chunked:
-        assert c.x.shape == (2, 4, 64, 64)
-        assert c.context.shape == (2, 77, 768)
-    # Last chunk's second batch row is the zero-pad.
-    assert torch.equal(chunked[-1].x[1], torch.zeros(4, 64, 64))
-
-
-def test_chunks_timestep_is_broadcast_from_first_value():
-    """t is rebuilt from t[0] across all chunks: locks current behavior that
-    discards any per-row timestep variation."""
-    x = torch.randn(2, 4, 64, 64)
-    t = torch.tensor([42.0, 99.0])  # the second value will be lost
-    context = torch.randn(2, 77, 768)
-    inputs = CoreMLInputs(x, t, context, control=None)
-    chunked = inputs.chunks(SD15_EXPECTED)
-    assert chunked[0].t.shape == (2,)
-    assert torch.equal(chunked[0].t, torch.full((2,), 42.0))

tests/unit/test_characterization_latents.py

@@ -1,118 +0,0 @@
-"""Characterization tests for coreml_suite.latents.
-
-Locks the *current* behavior of chunk_batch / merge_chunks — including the
-zero-pad regions and the truncation in merge — so a refactor
-cannot silently shift either contract.
-"""
-import pytest
-import torch
-
-from coreml_suite.core.latents import chunk_batch, merge_chunks
-
-
-@pytest.fixture(autouse=True)
-def _deterministic_seed():
-    torch.manual_seed(0)
-
-
-def _const_tensor(batch, *rest):
-    return torch.arange(batch * 4 * 8 * 8, dtype=torch.float32).reshape(batch, 4, 8, 8)
-
-
-# ---------- chunk_batch ------------------------------------------------------
-
-
-def test_chunk_batch_passthrough_when_shape_matches():
-    x = _const_tensor(2)
-    out = chunk_batch(x, (2, 4, 8, 8))
-    assert len(out) == 1
-    # passthrough: the same object identity is returned (no copy).
-    assert out[0] is x
-
-
-def test_chunk_batch_pads_single_chunk_when_input_smaller():
-    """batch=1, target=2 -> one padded chunk; the second row is exact zero."""
-    x = _const_tensor(1)
-    out = chunk_batch(x, (2, 4, 8, 8))
-    assert len(out) == 1
-    assert out[0].shape == (2, 4, 8, 8)
-    assert torch.equal(out[0][0], x[0])
-    assert torch.equal(out[0][1], torch.zeros(4, 8, 8))
-
-
-def test_chunk_batch_splits_exact_multiple():
-    """batch=4, target=2 -> two chunks, no padding."""
-    x = _const_tensor(4)
-    out = chunk_batch(x, (2, 4, 8, 8))
-    assert len(out) == 2
-    assert out[0].shape == (2, 4, 8, 8)
-    assert out[1].shape == (2, 4, 8, 8)
-    assert torch.equal(out[0], x[:2])
-    assert torch.equal(out[1], x[2:])
-
-
-def test_chunk_batch_pads_remainder_chunk():
-    """batch=5, target=2 -> chunks=[x[0:2], x[2:4]] then [x[4], 0]."""
-    x = _const_tensor(5)
-    out = chunk_batch(x, (2, 4, 8, 8))
-    assert len(out) == 3
-    assert torch.equal(out[0], x[0:2])
-    assert torch.equal(out[1], x[2:4])
-    last = out[-1]
-    assert last.shape == (2, 4, 8, 8)
-    assert torch.equal(last[0], x[4])
-    # The remainder row is zero-padded; lock that exact contract.
-    assert torch.equal(last[1], torch.zeros(4, 8, 8))
-    assert last[1].sum() == 0
-
-
-@pytest.mark.parametrize(
-    "batch_size,target,expected_chunks",
-    [
-        (1, 4, 1),
-        (3, 2, 2),
-        (5, 3, 2),
-        (9, 4, 3),
-    ],
-)
-def test_chunk_batch_pad_region_is_zero(batch_size, target, expected_chunks):
-    x = _const_tensor(batch_size)
-    out = chunk_batch(x, (target, 4, 8, 8))
-    assert len(out) == expected_chunks
-    mod = batch_size % target
-    if mod == 0 and batch_size >= target:
-        return
-    last = out[-1]
-    pad_rows = target - (mod if (mod != 0 and batch_size >= target) else batch_size)
-    pad_region = last[-pad_rows:]
-    assert torch.equal(pad_region, torch.zeros_like(pad_region))
-
-
-# ---------- merge_chunks -----------------------------------------------------
-
-
-def test_merge_chunks_exact_concat():
-    x = _const_tensor(4)
-    chunks = chunk_batch(x, (2, 4, 8, 8))
-    merged = merge_chunks(chunks, x.shape)
-    assert merged.shape == x.shape
-    assert torch.equal(merged, x)
-
-
-def test_merge_chunks_truncates_padding():
-    """Round-trip with a padded last chunk drops the pad rows."""
-    x = _const_tensor(5)
-    chunks = chunk_batch(x, (2, 4, 8, 8))
-    merged = merge_chunks(chunks, x.shape)
-    assert merged.shape == x.shape
-    assert torch.equal(merged, x)
-
-
-def test_merge_chunks_singleton_returns_equal_copy_when_shape_matches():
-    """A singleton chunk list still goes through torch.cat, so we get a new
-    tensor equal to the input — locked here because a refactor might be tempted
-    to short-circuit and accidentally return the same object."""
-    x = _const_tensor(2)
-    out = merge_chunks([x], x.shape)
-    assert torch.equal(out, x)
-    assert out is not x

tests/unit/test_characterization_sdxl_options.py

@@ -1,127 +0,0 @@
-"""Characterization tests for the SDXL options math.
-
-The SDXL time_ids / text_embeds math lives in
-coreml_suite.core.sdxl as pure builders. The framework adapter
-add_sdxl_model_options lives in models.py; here we just lock the pure
-math.
-"""
-import inspect
-
-import pytest
-import torch
-
-from coreml_suite.core.sdxl import (
-    build_sdxl_text_embeds,
-    build_sdxl_time_ids,
-    sdxl_model_function_wrapper,
-)
-
-
-@pytest.fixture(autouse=True)
-def _deterministic_seed():
-    torch.manual_seed(0)
-
-
-# ---------- build_sdxl_time_ids: base (len 6) -------------------------------
-
-
-def test_build_time_ids_base_defaults():
-    out = build_sdxl_time_ids({}, {}, is_base=True, is_refiner=False)
-    expected = torch.tensor([[768, 768, 0, 0, 768, 768], [768, 768, 0, 0, 768, 768]])
-    assert out.shape == (2, 6)
-    assert torch.equal(out, expected)
-
-
-def test_build_time_ids_base_respects_overrides():
-    pos = {"height": 1024, "width": 512, "crop_h": 8, "crop_w": 4,
-           "target_height": 1024, "target_width": 1024}
-    neg = {"height": 256, "width": 256, "crop_h": 0, "crop_w": 0,
-           "target_height": 256, "target_width": 256}
-    out = build_sdxl_time_ids(pos, neg, is_base=True, is_refiner=False)
-    expected = torch.tensor([[1024, 512, 8, 4, 1024, 1024], [256, 256, 0, 0, 256, 256]])
-    assert torch.equal(out, expected)
-
-
-# ---------- build_sdxl_time_ids: refiner (len 5) ----------------------------
-
-
-def test_build_time_ids_refiner_defaults():
-    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=True)
-    expected = torch.tensor([[768, 768, 0, 0, 6.0], [768, 768, 0, 0, 2.5]])
-    assert out.shape == (2, 5)
-    assert torch.equal(out, expected)
-
-
-def test_build_time_ids_refiner_respects_aesthetic_score():
-    pos = {"aesthetic_score": 8.5}
-    neg = {"aesthetic_score": 1.5}
-    out = build_sdxl_time_ids(pos, neg, is_base=False, is_refiner=True)
-    expected = torch.tensor([[768, 768, 0, 0, 8.5], [768, 768, 0, 0, 1.5]])
-    assert torch.equal(out, expected)
-
-
-# ---------- build_sdxl_time_ids: edge case ----------------------------------
-
-
-def test_build_time_ids_neither_base_nor_refiner_returns_len4():
-    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=False)
-    assert out.shape == (2, 4)
-
-
-# ---------- build_sdxl_text_embeds ------------------------------------------
-
-
-def test_text_embeds_concat_pos_then_neg():
-    pos = torch.full((1, 1280), 1.0)
-    neg = torch.full((1, 1280), -1.0)
-    out = build_sdxl_text_embeds(pos, neg)
-    assert out.shape == (2, 1280)
-    assert torch.equal(out[0], pos[0])
-    assert torch.equal(out[1], neg[0])
-
-
-# ---------- sdxl_model_function_wrapper closure -----------------------------
-
-
-def test_wrapper_captures_time_ids_text_embeds_refiner_via_closure():
-    time_ids = torch.zeros(2, 6)
-    text_embeds = torch.zeros(2, 1280)
-    wrapper = sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False)
-    closure = inspect.getclosurevars(wrapper).nonlocals
-    assert closure["time_ids"] is time_ids
-    assert closure["text_embeds"] is text_embeds
-    assert closure["refiner"] is False
-
-
-def test_wrapper_returns_zero_when_context_missing():
-    """When c_crossattn is None the wrapper short-circuits to zeros_like(x).
-    Locked here because the refactor mustn't change this default."""
-    wrapper = sdxl_model_function_wrapper(torch.zeros(2, 6), torch.zeros(2, 1280))
-    x = torch.randn(2, 4, 16, 16)
-    out = wrapper(
-        model_function=lambda *a, **kw: pytest.fail("model_function must not run"),
-        params={"input": x, "timestep": torch.zeros(2), "c": {}},
-    )
-    assert torch.equal(out, torch.zeros_like(x))
-
-
-def test_wrapper_refiner_truncates_context_to_g_clip():
-    """refiner=True slices c_crossattn[:, :, 768:] before forwarding."""
-    captured = {}
-
-    def fake_model(x, t, **c):
-        captured["context_shape"] = c["c_crossattn"].shape
-        captured["time_ids_shape"] = c["time_ids"].shape
-        return x
-
-    wrapper = sdxl_model_function_wrapper(
-        torch.zeros(2, 5), torch.zeros(2, 1280), refiner=True
-    )
-    x = torch.randn(2, 4, 16, 16)
-    context = torch.randn(2, 77, 2048)  # 768 + 1280 dims
-    wrapper(
-        model_function=fake_model,
-        params={"input": x, "timestep": torch.zeros(2), "c": {"c_crossattn": context}},
-    )
-    assert captured["context_shape"] == (2, 77, 1280)
-    assert captured["time_ids_shape"] == (2, 5)

tests/unit/test_chunks.py

@@ -1,34 +1,37 @@
-"""Smoke tests for the pure batch-chunking helpers in coreml_suite.core.
-
-Uses torch.device('cpu') instead of comfy.model_management.get_torch_device
-so Tier 0 runs without ComfyUI.
-"""
 import pytest
+
 import torch
 
-from coreml_suite.core.controlnet import chunk_control
-from coreml_suite.core.inputs import CoreMLInputs
-from coreml_suite.core.latents import chunk_batch, merge_chunks
-
-
-CPU = torch.device("cpu")
+from comfy.model_management import get_torch_device
+from coreml_suite.latents import chunk_batch, merge_chunks
+from coreml_suite.controlnet import chunk_control
+from coreml_suite.models import (
+    CoreMLInputs,
+)
+from coreml_suite.config import get_model_config
 
 
 @pytest.fixture
 def expected_inputs():
-    return {
+    expected = {
         "sample": {"shape": (2, 4, 64, 64)},
         "timestep": {"shape": (2,)},
         "timestep_cond": {"shape": (2, 256)},
-        "encoder_hidden_states": {"shape": (2, 77, 768)},
+        "encoder_hidden_states": {"shape": (2, 768, 1, 77)},
         "additional_residual_0": {"shape": (2, 320, 64, 64)},
         "additional_residual_1": {"shape": (2, 640, 32, 32)},
     }
+    return expected
+
+
+@pytest.fixture
+def model_config():
+    return get_model_config()
 
 
 @pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
 def test_batch_chunking(batch_size):
-    latent_image = torch.randn(batch_size, 4, 64, 64).to(CPU)
+    latent_image = torch.randn(batch_size, 4, 64, 64).to(get_torch_device())
     target_shape = (4, 4, 64, 64)
 
     chunked = chunk_batch(latent_image, target_shape)
@@ -42,7 +45,7 @@ def test_batch_chunking(batch_size):
 
 @pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
 def test_merge_chunks(batch_size):
-    input_tensor = torch.randn(batch_size, 4, 64, 64).to(CPU)
+    input_tensor = torch.randn(batch_size, 4, 64, 64).to(get_torch_device())
     target_shape = (4, 4, 64, 64)
     chunked = chunk_batch(input_tensor, target_shape)
 
@@ -54,16 +57,16 @@ def test_merge_chunks(batch_size):
 
 @pytest.fixture
 def inputs():
-    x = torch.randn(1, 4, 64, 64).to(CPU)
-    t = torch.randn([1]).to(CPU)
-    c_crossattn = torch.randn(1, 77, 768).to(CPU)
+    x = torch.randn(1, 4, 64, 64).to(get_torch_device())
+    t = torch.randn([1]).to(get_torch_device())
+    c_crossattn = torch.randn(1, 77, 768).to(get_torch_device())
     control = {
         "output": [
-            torch.randn(1, 320, 64, 64).to(CPU),
-            torch.randn(1, 640, 32, 32).to(CPU),
+            torch.randn(1, 320, 64, 64).to(get_torch_device()),
+            torch.randn(1, 640, 32, 32).to(get_torch_device()),
         ],
     }
-    timestep_cond = torch.randn(1, 256).to(CPU)
+    timestep_cond = torch.randn(1, 256).to(get_torch_device())
 
     return CoreMLInputs(x, t, c_crossattn, control, timestep_cond=timestep_cond)
 
@@ -83,11 +86,11 @@ def inputs():
 def test_chunking_controlnet(b, target_size, num_chunks):
     cn = {
         "output": [
-            torch.randn(b, 320, 64, 64).to(CPU),
-            torch.randn(b, 640, 32, 32).to(CPU),
+            torch.randn(b, 320, 64, 64).to(get_torch_device()),
+            torch.randn(b, 640, 32, 32).to(get_torch_device()),
         ],
         "middle": [
-            torch.randn(b, 1280, 8, 8).to(CPU),
+            torch.randn(b, 1280, 8, 8).to(get_torch_device()),
         ],
     }
 

tests/unit/test_tier0_purity.py

@@ -1,43 +0,0 @@
-"""Gate: prove the Tier-0 lane is framework-free.
-
-In a pure `pytest -m unit` run, none of the banned runtime modules
-(comfy, coremltools, python_coreml_stable_diffusion, folder_paths,
-nodes, comfy_extras, diffusers, diffusionkit) may be in sys.modules
-after collection. If they are, a tests/unit/ file is transitively
-pulling them in and the Tier-0 promise — "runs on Linux with no Mac
-stack" — is broken.
-
-When other tiers are also collected, framework modules may be imported
-deliberately (e.g. smoke pulls in coremltools), so the check is skipped
-unless the run is purely `-m unit` — Tier-0 purity is only meaningful
-when nothing else is loaded.
-"""
-import sys
-
-import pytest
-
-BANNED_ROOTS = {
-    "comfy",
-    "comfy_extras",
-    "coremltools",
-    "python_coreml_stable_diffusion",
-    "folder_paths",
-    "nodes",
-    "diffusers",
-    "diffusionkit",
-}
-
-
-def test_no_framework_modules_loaded_by_unit_tier(request):
-    markexpr = request.config.option.markexpr
-    if markexpr != "unit":
-        pytest.skip(
-            "purity gate only meaningful in a pure `-m unit` run "
-            f"(got markexpr={markexpr!r}); other tiers are expected to "
-            "import comfy/coremltools."
-        )
-    loaded = {name for name in sys.modules if name.split(".")[0] in BANNED_ROOTS}
-    assert not loaded, (
-        f"Tier-0 leakage: these framework modules are in sys.modules after "
-        f"collecting tests/unit/: {sorted(loaded)}. Pure-core promise broken."
-    )