The native integration removes the separate custom-node installation used by earlier community workflows. The official template includes both models, with a switch selecting TRELLIS.2. Geometry generation is only the first part: mesh cleanup and material baking make the output easier to take into a 3D editor.
What the native workflow generates
TRELLIS.2 uses a 4B-parameter image-to-3D model and O-Voxel structured latents. The pipeline builds sparse structure, refines shape, and generates base color, metallic, and roughness materials.
The native workflow adds remeshing, face-count reduction, UV unwrapping, and normal/AO baking. These steps are useful when a detailed raw mesh needs a smaller, textured deliverable.
Load the template and choose TRELLIS.2
Update ComfyUI and open the template named “Pixal3D & TRELLIS.2: Image to Model”. Follow its model-download notes, then load a clear object image in LoadImage.
Set Boolean (Switch to Trellis2) to true. The shared template defaults to Pixal3D, so leaving this switch unchanged runs a different model. Queue the workflow and inspect the result in Preview3DAdvanced; the official guide saves it under ComfyUI/output/3d/ComfyUI/.
Put each checkpoint in the correct folder
The template requires both diffusion checkpoints because it can switch between the two pipelines. Shape and texture VAEs, DINOv3 image conditioning, and BiRefNet background removal are shared components.
Use the filenames and folders shown below. Missing nodes usually point to an outdated ComfyUI build or a failed node import; a missing checkpoint is a separate model-path problem.
ComfyUI/models/
├── diffusion_models/
│ ├── trellis_2_int8_convrot.safetensors
│ └── pixal3d_int8_convrot.safetensors
├── vae/
│ ├── trellis_2_shape_vae_bf16.safetensors
│ └── trellis_2_texture_vae_bf16.safetensors
├── clip_vision/
│ └── dino_v3_L_naf_fp32.safetensors
└── background_removal/
└── birefnet.safetensorsInspect the mesh and materials before export
Check the silhouette, thin parts, holes, and face count before choosing the final asset. Remeshing and decimation solve different problems: one rebuilds surface connectivity, while the other reduces the mesh budget.
The model generates base color, metallic, and roughness. Normal and ambient-occlusion maps are baked from the high-poly geometry in the ComfyUI implementation. Inspect them under neutral lighting in your 3D editor instead of judging only a bright preview.
TRELLIS.2 and Pixal3D in the shared template
| Setting | TRELLIS.2 | Pixal3D |
|---|---|---|
| Template switch | Boolean (Switch to Trellis2) = true | Default pipeline when switch is false |
| Diffusion checkpoint | trellis_2_int8_convrot.safetensors | pixal3d_int8_convrot.safetensors |
| Image conditioning | TRELLIS.2 conditioning | Pixel-aligned conditioning |
| Shared stages | Shape/texture VAEs and mesh processing | Shape/texture VAEs and mesh processing |
Frequently asked questions
Do I still need a TRELLIS.2 custom-node pack?
Not for this native workflow. The models use core ComfyUI nodes. Older third-party workflows can still reference their own custom nodes, so use the current official template.
How much GPU memory do I need?
Microsoft's reference implementation requires at least 24GB of NVIDIA GPU memory. In ComfyUI, resolution, precision, and model offloading also affect memory use. If you run out of memory, lower the generation resolution and check the memory-management settings.
Why are nodes or models missing?
Update ComfyUI first. Check startup errors for failed imports, then verify that each model is in the matching folder. The official docs note that some newly documented core nodes may require a Nightly build before the next stable release.
Are normal and AO maps predicted by TRELLIS.2?
The native workflow bakes those maps from the high-poly mesh. TRELLIS.2's material model generates base color, metallic, and roughness; the baking stage extends the final texture set.