Trellis 2: ComfyUI image-to-3D workflow and setup

This tutorial covers a Trellis 2 image-to-3D workflow in ComfyUI and its installation on a portable Windows build. RMBG 2.0 removes the background before Trellis generates geometry. Mesh repair and refinement precede texturing. The speaker selects Microsoft Trellis 24B and recommends SDPA when Flash Attention is unavailable.

The examples show how source images affect results: a dark, blurry car image produces weak textures, while a complex shape takes longer to generate. The speaker reports roughly 328 seconds for one generation and 6 to 9 GB of VRAM use. They suggest the 1024 cascade for an 8 GB card and 1536 for a 12 GB card. Reported system RAM use reaches around 32 GB for simpler models and fills 64 GB with a complex example at the maximum pipeline setting.

Setup covers Git cloning, five wheel files matched to Python 3.13, dependency repair through ComfyUI Manager, and a Hugging Face read token stored as HF_TOKEN. The walkthrough downloads complete DINOv2 and RMBG 2.0 repositories into model folders. The speaker says DINOv2 requires an access agreement.

The build versions conflict: the description names Torch 2.10 and CUDA 13, while the transcript says Torch 2.1.0 and CUDA 12.1. Generated meshes can have gaps, and textures may need work in specialized software.