commit f529e76ff666876cf888c8be62dffde2b38d440a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Jul 10 13:48:21 2026 +0200

media-gfx/comfyui: drop 0.26.1

commit 588ca3cb0238eca4dbab22fec5180336d2d26c4c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Jul 10 13:47:15 2026 +0200

media-gfx/comfyui: add 0.27.1

commit f085293b1efac43ee382bd820357b39230c339ec
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Jul 5 14:30:48 2026 +0200

media-gfx/comfyui: condense build-notes comments

commit c2c1de2d11c5457408696c83c7e61dca7501b5b3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:39:47 2026 +0200

media-gfx/comfyui: drop 0.25.1, 0.26.0

commit 84220a6e503437d52b32a1a0e45bddec508a3f39
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 1 09:39:47 2026 +0200

media-gfx/comfyui: add 0.27.0

commit 7e1f36c37be17ef373e8e13da0567599e1f3f371
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Jun 26 13:30:27 2026 +0200

media-gfx/comfyui: add 0.26.2

commit 63b7ac7d389b8267b085aba7be5d378839503120
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 25 16:51:35 2026 +0200

media-gfx/comfyui: add 0.26.1

commit 1fd181e0e253e7f1622abf1d211593d35b62e498
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Jun 23 22:43:26 2026 +0200

media-gfx/comfyui: drop 0.24.1, 0.25.0

commit b38ac6827858b9dddae8d7ee928605513a384f11
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Jun 23 22:42:57 2026 +0200

media-gfx/comfyui: add 0.26.0

commit 2d045c63e9461e9682164100509e843589a731fc
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 08:02:02 2026 +0200

media-gfx/comfyui: add 0.25.1

commit eab940beae0eff093e28aa1631bd4915c44d604c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Jun 16 10:48:38 2026 +0200

media-gfx/comfyui: add 0.25.0

commit 9150db2fca551a87954b6fee54d78070ba1fea6c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 13:56:14 2026 +0200

media-gfx/comfyui: gate torchaudio behind audio USE flag

torchaudio was an unconditional RDEPEND, but it tops out at 2.11.0
upstream (pins ~sci-ml/pytorch-2.11) while the stack is pytorch-2.12 /
caffe2-2.12. emerge --deep comfyui therefore backtracked into downgrading
pytorch to 2.11, which then collides with caffe2-2.12 -- the whole stack
was unresolvable on a fresh pull (a non-deep resolve hid it because
torchaudio was already installed).

comfyui imports torchaudio only in the audio-node paths
(comfy.audio_encoders, the lumina audio VAE), lazily and try/except-
wrapped, so image generation never touches it. Move it behind a
default-off USE=audio: the default stack now resolves, and audio is an
honest opt-in for once a torchaudio matching the installed torch exists.

commit d66c7840cafc5b63171591bc78ee81bcfe8fc54c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 10:23:17 2026 +0200

media-gfx/comfyui: add rocm USE flag

caffe2 builds with either cuda or rocm acceleration, but comfyui only
exposed a cuda flag, leaving AMD users on the CPU fallback. Add a rocm
flag, mutually exclusive with cuda via ?? ( cuda rocm ), that builds the
torch backend as caffe2[rocm].

comfy_aimdo / comfy_kitchen ship CUDA-only wheels, so on the AMD path
they install their py3-none-any builds (cuda= off). That leaves
comfy_kitchen.apply_rope -- which the flux/lumina/z-image families call
unconditionally at sampling -- without a kernel; dev-python/triton-bin
supplies it through its AMD Triton backend, so rocm pulls triton-bin the
same way compile already does. comfy_aimdo's VRAM offload is a no-op
under ROCm but blocks nothing.

commit 3ec5c9daecebf498c18c3b3190ed892a4b412ebb
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 00:16:02 2026 +0200

media-gfx/comfyui: add 0.24.1

Node-based diffusion / generative-AI workflow GUI, run from source under
/opt/comfyui with a /usr/bin/comfyui launcher that redirects per-user state
(models, custom_nodes, input, output, user, temp) to --base-directory,
defaulting to ~/.local/share/comfyui. The launcher pre-creates those subdirs
(ComfyUI's prestartup lists custom_nodes without creating it) and points
--database-url at <base>/user/comfyui.db, since the default DB path lives
under the read-only /opt tree. A system-wide custom_nodes dir at
$/usr/share/comfyui/custom_nodes (kept via keepdir and wired through a
generated extra_model_paths.yaml) lets admin- or emerge-installed nodes sit
alongside the per-user <base>/custom_nodes.

python-single-r1 with mixed USE-deps: the single-impl torch/HF stack via
PYTHON_SINGLE_USEDEP, multi-impl libs via python_gen_cond_dep. comfy_aimdo
(VRAM allocator) is a hard module-load import, and comfy_kitchen supplies the
RoPE/quant kernels the flux/lumina/z-image families use unconditionally via
comfy.quant_ops.ck -- so dev-python/comfy-aimdo-bin and comfy-kitchen-bin are
both mandatory. USE=cuda selects their CUDA wheels (+ caffe2[cuda], propagated
via cuda=); without cuda their py3-none-any fallbacks + a CPU caffe2 give a
slow CPU path. USE=compile pulls dev-python/triton-bin for torch.compile +
comfy_kitchen's Triton backend. gguf in upstream's freeze is the ComfyUI-GGUF
custom node, not core, and is excluded.

USE: cuda, compile, templates (the ~330 MiB example-gallery meta), extra
(spandrel + kornia), opengl (pyopengl + glfw). The comfyui-* data packages are
pinned to the versions 0.24.1's requirements.txt freezes.