Trellis2 Apple MLX Image to 3D (Advanced)

This is the exact same node as Trellis2AppleGenerate, with six extra dials bolted on. Same pipeline input, same image, same seed, same pipeline_type choices, same mesh output - but now you can steer the sampler directly instead of taking the author's smoke-tested defaults on faith. The README is blunt about when you should use it: only when you want to tune guidance and steps manually. Otherwise the simple node is the better choice, because its defaults are the ones that have actually been run.

Why six dials and not one

TRELLIS.2's MLX backend generates in three stages, and each gets its own guidance/steps pair:

• ss_* - the sparse-structure stage. This decides the coarse 3D layout, the rough shape of the volume. Get this wrong and nothing downstream can save it.
• shape_* - the shape-slat stage, where actual geometry forms. This is the one that matters most for how the model looks.
• tex_* - the texture stage, painting the surface. Default guidance is 1.0 because texturing is a much easier job than geometry - it doesn't need the same shove.

Every guidance field runs 0.0–20.0 in 0.1 steps; every steps field runs 1–50. The defaults (7.5 / 7.5 / 1.0, 12 / 12 / 12) match the simple node exactly, so you can drop this in as a drop-in replacement and nothing changes until you move a dial.

What's actually worth touching

Honest guidance for a beginner:

• Steps past ~20 mostly buy time, not quality. These are flow-based samplers; there's a point of diminishing returns fast. If you're not happy at 12, try 20 before you try 50.
• Lower guidance if the mesh comes out overcooked. Too-high guidance on flow models tends to mean artifacts, oversaturated texture, or geometry that's fighting itself. The Advanced node exists so you can dial shape_guidance down to 5 or 6 and see.
• Crank ss_steps if the structure is unstable. Wobbly silhouettes and floating bits usually point at the sparse-structure stage, not the texture stage. Add steps there first.
• Leave tex_guidance near 1. It's low for a reason; pushing texture guidance hard is where you get painted-on-looking surfaces.

Setup and gotchas

Install is the same shared pack ritual - clone pedronaugusto/trellis2-apple, symlink this repo into custom_nodes/, run install.py with ComfyUI's own Python plus --download-weights, then doctor.py:

git clone https://github.com/pedronaugusto/trellis2-apple /path/to/trellis2-apple
ln -s /path/to/ComfyUI-Trellis2Apple /path/to/ComfyUI/custom_nodes/ComfyUI-Trellis2Apple
/path/to/ComfyUI/.venv/bin/python custom_nodes/ComfyUI-Trellis2Apple/install.py \
  --trellis2-apple /path/to/trellis2-apple --download-weights

The same caveats apply here as everywhere in this pack: Apple Silicon only, a 4B model that wants real unified memory (the Mac port family's comfortable floor is around 24 GB), and transformers<5 pinned by the backend, which can fight other nodes in the same ComfyUI env.

Tuning is iterative and each run is minutes, not seconds - so change one thing at a time, keep the seed fixed so you're actually comparing apples to apples, and only reach for this node when the simple defaults have a specific complaint you're trying to fix. Most of the time, Trellis2AppleGenerate with 512 - fast is the right call, and this node is the toolbox you open when it isn't.