Trellis 2 and Pixal3D: native ComfyUI setup

This tutorial covers native Trellis 2 and Pixal3D workflows in ComfyUI for local AI image-to-3D generation. The presenter credits Kijai with rebuilding the model code and post-processing in pure PyTorch without extra dependencies. Setup uses ComfyUI Desktop or Portable; the recording requires the latest GitHub version rather than the stable release.

The walkthrough covers downloading model files, placing them in the appropriate folders, and selecting them in red nodes after a restart or refresh. The presenter recommends Pixal3D Int8 for 6–8GB VRAM and also demonstrates BF16. He specifies a 6GB minimum for quantized workflows. AMD and Intel GPUs were not tested by his team, so compatibility remains unconfirmed.

Switching between the two workflows requires both the matching model and conditioning node. Quality adjustments include raising sampling steps above the default 12, changing texture resolution, and reducing the mesh face target from 700,000 to around 300,000. The presenter notes that further retopology and texture baking are still needed.

Limitations include no native multi-view support and possible VRAM spikes during 4K UV unwrapping on smaller cards. Practical tips cover unloading models and the execution cache, manually exporting GLB when automatic saving fails, and disabling texture baking when only geometry is needed. Blender appears as a way to inspect the exported mesh, textures, and wireframe.