Skip to main content
TRELLIS.2 is an open-source image-to-3D model from Microsoft Research, and the successor to TRELLIS. It is a 4B-parameter model built on native 3D VAEs with 16× spatial compression and a sparse voxel representation called O-Voxel, producing detailed 3D assets with full PBR materials from a single image. TRELLIS.2 shares its base architecture with Pixal3D: both use the same DINOv3 image encoder, the same shape and texture VAEs, and the same structure, shape, and texture generation stages. Only the main diffusion model and the conditioning format differ. If you have run the Pixal3D workflow, the model setup is almost identical.

​
TRELLIS.2: Image to Model

Upload a single image. Generate a high-fidelity 3D model with full PBR textures using TRELLIS.2. This workflow also includes the Pixal3D pipeline. To run TRELLIS.2, set the Boolean (Switch to Trellis2) node to true. The workflow automatically switches to the TRELLIS.2 model and its conditioning. TRELLIS.2 workflow preview
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

Run on Comfy Cloud

Run this workflow instantly on Comfy Cloud

Download Workflow

Download JSON or search “Pixal3D & TRELLIS.2: Image to Model” in Template Library
Input materials Upload this file to the LoadImage node:

viking_wolf_rune_axe.png

LoadImage node 122 · viking_wolf_rune_axe.png
Input image

​
How it works

TRELLIS.2 generates a 3D asset in three diffusion stages and then runs a full mesh post-processing pipeline:
  1. Background removal: BiRefNet removes the background from the input image, and the workflow crops the subject to a centered canvas. A switch lets you skip background removal
  2. Structure generation: a sparse structure latent is sampled and decoded into voxels, then converted into a rough mesh
  3. Shape refinement: the shape stage and the upsampling stage refine the mesh up to the target resolution (1536)
  4. Texture generation: a texture diffusion stage produces PBR material voxels (base color, metallic, roughness)
  5. Post-processing: DC remesh, QEM decimation, UV unwrapping, and baking of base color, normal, and ambient occlusion maps into the final textured mesh

​
Steps to run

  1. Load an image: use the LoadImage node to load a single image of the object
  2. Switch to the TRELLIS.2 pipeline: set Boolean (Switch to Trellis2) to true. The workflow loads the TRELLIS.2 checkpoint automatically
  3. Queue the workflow: press Ctrl (Cmd on macOS) + Enter
  4. Wait for the pipeline: the structure, shape, and texture stages run in sequence, followed by post-processing
  5. View the result: inspect the mesh in the Preview3DAdvanced node. The 3D model is saved to ComfyUI/output/3d/ComfyUI/

​
Model downloads

Download the models used by this workflow. Both diffusion checkpoints are required: the switch selects which one runs. Place them in the corresponding models/ subdirectories.

TRELLIS.2 UNet

trellis_2_int8_convrot.safetensors: TRELLIS.2 diffusion model, loaded when the switch is set to true

Pixal3D UNet

pixal3d_int8_convrot.safetensors: Pixal3D diffusion model, loaded by default

Shape VAE

trellis_2_shape_vae_bf16.safetensors: VAE for structure and shape decoding

Texture VAE

trellis_2_texture_vae_bf16.safetensors: VAE for texture decoding

DINOv3 CLIP vision

dino_v3_L_naf_fp32.safetensors: CLIP vision encoder for image conditioning. The equivalent dino_v3_vit_l.safetensors from the TRELLIS.2 repo also works

BiRefNet background removal

birefnet.safetensors: background removal model for preprocessing

​
Model storage location