Trellis 2 vs Pixal3D: ComfyUI setup and 3D rendering

This tutorial covers image-to-3D workflows in ComfyUI with Trellis 2 and Pixal3D, followed by image rendering with Flux Klein. Setup includes updating ComfyUI and Pixaroma Nodes, downloading the required models, and selecting their locations in the workflow. The presenter explains a manual update route because the node manager showed an older release at recording time.

The first Trellis 2 workflow crops the input to a square, removes its background, generates geometry, adds colors, and exports a GLB file. A Blender import demonstrates how to inspect the result in another 3D application. A second workflow adds textures and uses a higher face count.

For local AI hardware, the presenter reports testing on 6 GB and 24 GB VRAM cards. The basic Trellis workflow ran on 6 GB, while its textured variant could take around 40 minutes. The textured Pixal3D run exceeded two hours on that card before the presenter stopped it. In these tests, Pixal3D preserved the input camera angle but sometimes invented unwanted geometry on the back; Trellis was faster.

The final workflows turn a positioned 3D model into an image with Flux Klein. The presenter shows framing problems with close-ups and nonsquare images in one workflow, then uses another to retain composition. Suggested starting denoise values are 0.7 to 0.8, with results also depending on the prompt and seed.