media-gfx/comfyui-if-trellis
ComfyUI custom node: TRELLIS image-to-3D (Gaussian/.ply generation)
Screenshots
View
Download
BrowseChangeLog
commit e62db9379215cd6865a3d46810b8f2dd95d33808
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 12:41:32 2026 +0200
media-gfx/comfyui-if-trellis: correct the commit-pinning rationale
The comment said upstream tags no releases. It does: 0.0.2, 0.1.0 and 0.1.1
all exist. What makes them unusable is that the newest, 0.1.1, points at a
commit from 2025-01-06, two months behind the 2025-03-09 commit we pin.
The conclusion is unchanged - pinning the commit is still right - but the
stated reason was wrong, and a future bump checking "are there tags yet?"
would have found some and drawn the wrong conclusion.
commit 1ce295bf69d2099658c81db25868845bdce145c1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 12:41:03 2026 +0200
media-gfx/comfyui-if-trellis: add the missing imageio-ffmpeg dependency
imageio-ffmpeg is an unconditional entry in the pinned commit's pyproject
dependencies and requirements.txt, and IF_Trellis.py writes its preview as
an .mp4 through imageio.mimsave(). imageio on its own cannot encode that -
the ffmpeg plugin is a separate distribution.
Declared unconditionally rather than behind a USE flag because the widget
that reaches the call site, render_video, is a runtime toggle in the node's
inputs, not something resolvable at build time.
commit f0e2e793a8caec9df5bde24b15fda6a7fb8a287f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 14:36:32 2026 +0200
media-gfx/comfyui-if-trellis: new package, add 0_p20250309
ComfyUI custom node wrapping Microsoft's TRELLIS single-image-to-3D model.
Pins upstream commit 51bdfc6 (no tagged releases) and ships the bundled
trellis/ library; the prebuilt wheels/ and the extensions/ source tree are
excluded in favour of our own from-source CUDA-extension ebuilds.
The sparse-conv backend is spconv, not torchsparse: the TRELLIS pretrained
weights are stored in spconv's weight layout, which torchsparse cannot load,
so RDEPEND pulls dev-python/spconv-cu126.
Four local patches keep the node usable with the deps we package: 0001 defers
rembg to call time (RGBA inputs work without it); 0002 makes the mesh
post-processing imports optional so the Gaussian .ply path loads cleanly; 0003
stubs the optional kaolin FlexiCubes asserts; 0004 guards an unconditional
texture_image.shape log on the gaussian-only path.
USE=mesh pulls the .glb export stack (xatlas / pyvista / python-igraph /
pymeshfix). The Gaussian (.ply) path works without it.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 12:41:32 2026 +0200
media-gfx/comfyui-if-trellis: correct the commit-pinning rationale
The comment said upstream tags no releases. It does: 0.0.2, 0.1.0 and 0.1.1
all exist. What makes them unusable is that the newest, 0.1.1, points at a
commit from 2025-01-06, two months behind the 2025-03-09 commit we pin.
The conclusion is unchanged - pinning the commit is still right - but the
stated reason was wrong, and a future bump checking "are there tags yet?"
would have found some and drawn the wrong conclusion.
commit 1ce295bf69d2099658c81db25868845bdce145c1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 12:41:03 2026 +0200
media-gfx/comfyui-if-trellis: add the missing imageio-ffmpeg dependency
imageio-ffmpeg is an unconditional entry in the pinned commit's pyproject
dependencies and requirements.txt, and IF_Trellis.py writes its preview as
an .mp4 through imageio.mimsave(). imageio on its own cannot encode that -
the ffmpeg plugin is a separate distribution.
Declared unconditionally rather than behind a USE flag because the widget
that reaches the call site, render_video, is a runtime toggle in the node's
inputs, not something resolvable at build time.
commit f0e2e793a8caec9df5bde24b15fda6a7fb8a287f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 14:36:32 2026 +0200
media-gfx/comfyui-if-trellis: new package, add 0_p20250309
ComfyUI custom node wrapping Microsoft's TRELLIS single-image-to-3D model.
Pins upstream commit 51bdfc6 (no tagged releases) and ships the bundled
trellis/ library; the prebuilt wheels/ and the extensions/ source tree are
excluded in favour of our own from-source CUDA-extension ebuilds.
The sparse-conv backend is spconv, not torchsparse: the TRELLIS pretrained
weights are stored in spconv's weight layout, which torchsparse cannot load,
so RDEPEND pulls dev-python/spconv-cu126.
Four local patches keep the node usable with the deps we package: 0001 defers
rembg to call time (RGBA inputs work without it); 0002 makes the mesh
post-processing imports optional so the Gaussian .ply path loads cleanly; 0003
stubs the optional kaolin FlexiCubes asserts; 0004 guards an unconditional
texture_image.shape log on the gaussian-only path.
USE=mesh pulls the .glb export stack (xatlas / pyvista / python-igraph /
pymeshfix). The Gaussian (.ply) path works without it.