Image to 3D Model

Trellis.2 is Microsoft's open-source image-to-3D model, released in December 2025. It is a 4-billion-parameter model built on a compact O-Voxel structured latent representation that generates geometry and PBR materials from a single image at effective resolutions up to 1536³. It now runs natively in ComfyUI core.

Pixal3D is an image-to-3D model from Tsinghua University and Tencent ARC Lab. It uses pixel-aligned generation to keep the 3D output in direct correspondence with the input view, producing near-reconstruction-level fidelity. It is built on the Trellis.2 backbone and runs natively in ComfyUI alongside it.

Yes. Trellis.2's code and weights are MIT-licensed, and ComfyUI's native implementation replaces the non-commercial nvdiffrast and nvdiffrec dependencies from the original pipeline with a from-scratch PyTorch and SciPy reimplementation. There are no non-commercial dependencies in the ComfyUI pipeline, so generated assets can be used in commercial work.

Pixal3D shares the Trellis.2 backbone, VAEs, and DINOv3 image conditioning, and adds pixel-aligned generation, explicit image back-projection into a 3D feature volume, and cascaded refinement of structure, shape, and texture. It generally produces better results than Trellis.2, with tighter fidelity to the input image. Both output the same PBR material set.

Nothing per asset. Both models are open weights and can run locally on consumer hardware as part of ComfyUI core, so there is no per-generation charge. They are also available on Comfy Cloud, where standard Comfy Cloud credit usage applies.

No. The original Trellis.2 pipeline required PyTorch 2.6.0 with CUDA 12.4 and a stack of compiled extensions. The native ComfyUI implementation drops all of them. If your ComfyUI runs, these models run on your current PyTorch.

A single image. Both models generate geometry and materials from one input view. The workflows also include BiRefNet for background removal and MoGe for camera field-of-view estimation, both available on the Comfy-Org Hugging Face page.

Comfy-Org/TRELLIS.2 and Comfy-Org/Pixal3D from Hugging Face, plus Comfy-Org/BiRefNet for background removal and Comfy-Org/MoGe for camera FOV estimation. The workflow note lists the target directory for each.

Trellis.2's texturing model produces base color, roughness, and metallic maps. The ComfyUI implementation adds a UV unwrapping node and bakes two more maps from the high-poly mesh: a normal map and an ambient occlusion map, giving a complete PBR material set.

Four new nodes ship with this release. Remesh Mesh fixes holes and mesh imperfections, Decimate Mesh reduces face and vertex count to a target budget, Smooth Mesh Normals smooths the mesh volume, and Render Mesh produces a quick image render. The Load 3D, Preview 3D, and Save 3D nodes were also rebuilt for these workflows.