Bundles all post-v0.1.2 polish into a single release commit.
Code
----
* nodes/pixal3d_stages.py: short-circuit `_rescue_inference_tensors`
when the caller has already disabled `inference_mode`. `run_pixal3d`
always wraps the pipeline call in `torch.inference_mode(mode=False)`,
so every tensor allocated inside is already a regular tensor and the
rescue is a no-op — but iterating 4× DinoV3 ViT-L + Pixal3D parameters
at Python loop speed was burning ~20 min single-threaded on cold-load.
One-line guard now skips that.
Packaging
---------
* pyproject.toml: 0.1.2 → 0.1.3 — re-triggers publish-comfy-registry.yml.
Workflows (both basic + external-rembg variants)
------------------------------------------------
* Trimmed to the minimal useful graph: removed the 8-view
Trellis2RenderPreview node and its PreviewImage downstream, removed
the ShowText|pysssss "GLB always saved to output/…" note (now
redundant with the in-graph export), removed the colored group
bounding box.
* Added Trellis2ExportTrimesh (TRIMESH → STRING path) → Preview3D
(built-in interactive 3D viewer) so users see the textured mesh
rotated in-canvas on every queue.
* Three-lane integer-aligned grid layout: Inputs (LoadImage + optional
external rembg chain) → Pipeline (Pixal3DImageToMesh sized 380×950
so every widget is visible without scroll) → Outputs (preprocessed
preview, Save GLB, 3D viewer). Viewport `extra.ds` pre-framed for
1080p first-open.
* In-canvas Note node with the verified low-VRAM tweak ladder for
12–16 GB cards.
* Pointed LoadImage at the new demo subject (examples/test_image.png)
and bumped defaults to a verified-safe high-fidelity combo:
1024_cascade + 16/16/16 sampling steps + 65k tokens + 300k decimation
+ 4096 texture. ~3 min wall time when warm; ~7–10 min cold. Does
NOT trip the comfy-aimdo 16 GB cast-buffer + 1536_cascade-models
OOM ceiling at ~46 GB.
Docs / demo images
------------------
* docs/images/demo_4view.png — four PBR-textured views of the new
subject (front, right-3/4, back, left-3/4) rendered offline at 2×
SSAA via pyrender. Bypasses Trellis2RenderPreview's spike/aliasing
artifacts entirely.
* docs/images/demo_mesh_frontal.png — clay-shaded frontal of the same
mesh so reviewers can read geometry quality independent of the PBR
bake. Lighting tuned for mid-tone exposure without burnout.
* docs/images/nodes_preview.png — real ComfyUI canvas screenshot of
the v0.1.3 workflow (not a render of the old workflow).
* examples/test_image.png — new demo input (baroque/skull/harp).
README
------
* Hero block updated with both demo images + clarifying captions
(textured + clay).
* VRAM-requirements line now points at the workflow's in-canvas Note
for low-VRAM tuning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
