TRELLIS.2 and Pixal3D Image-to-3D Generation Now Built into ComfyUI

ComfyUI Wikinews

ComfyUI core now supports TRELLIS.2 and Pixal3D image-to-3D: shape and texture stages, MoGe depth, and mesh post-processing nodes with official Comfy-Org weights.

TRELLIS.2 (Hugging Face | GitHub | Project Page) and Pixal3D (Hugging Face | GitHub) image-to-3D generation are now built directly into ComfyUI core: the integration PR (Comfy-Org/ComfyUI #14718) merged on August 22, 2026, so no custom node pack is required.

Overview

TRELLIS.2 is Microsoft's 4-billion-parameter image-to-3D model. It uses the O-Voxel (omni-voxel) sparse voxel representation with a flow-matching transformer to generate high-fidelity 3D assets with complex topologies, sharp features, and full PBR materials including transparency. The model was released in December 2025 (see our previous coverage) and requires no rendering or optimization passes during conversion.

Pixal3D is a SIGGRAPH 2026 model from Tencent ARC Lab and Tsinghua University, built on the TRELLIS.2 backbone. It uses pixel-aligned features via back-projection to establish direct pixel-to-3D correspondences, reaching near-reconstruction-level fidelity with detailed geometry and PBR textures from a single image.

TRELLIS.2 texturing example: The Forgotten Knight

TRELLIS.2 texture-stage example (source: microsoft/TRELLIS.2)

Pixal3D image-to-3D results

Pixal3D single-image reconstruction results (source: Pixal3D project page)

Native ComfyUI Support

The integration PR adds a dependency-free implementation of both models plus a complete mesh post-processing chain, all re-implemented in PyTorch/scipy with no extra dependencies:

  • TRELLIS.2 nodes — Trellis2ShapeStage, Trellis2TextureStage, Trellis2UpsampleStage, Trellis2Conditioning, EmptyTrellis2LatentStructure, plus dedicated shape/texture VAE decoders (VaeDecodeShapeTrellis, VaeDecodeTextureTrellis)
  • Pixal3D nodes — Pixal3DConditioning alongside the shared TRELLIS.2 stages; both models share the same VAEs and DINOv3 image encoder
  • MoGe metric depth nodes — LoadMoGeModel, MoGeInference, MoGePanoramaInference, MoGePointMapToMesh, MoGeGeometryToFOV, MoGeRender for camera/FoV estimation and geometry conditioning
  • Mesh post-processing — RemeshMesh (DC remesh), DecimateMesh (QEM decimation), UnwrapMesh (UV unwrap), ApplyTextureToMesh, BakeTextureFromVoxel, BakeAmbientOcclusion, BakeNormalMapFromMesh, FillHoles, WeldVertices, MeshSmoothNormals
  • 3D export — SaveGLB, Save3DAdvanced, SaveGaussianSplat, SavePointCloud, MeshToFile3D, plus MergeMeshes and RotateMesh utilities
TRELLIS.2 / Pixal3D workflow inside ComfyUI

Test workflow for Pixal3D inside ComfyUI (source: Comfy-Org/ComfyUI PR #14718)

Workflow

The official example workflow, hosted in the Comfy-Org/Pixal3D repository, runs the full chain from a single input image to shape generation, texture baking, and mesh export:

Model Download

Comfy-Org repackaged weights provide ready-to-use checkpoints in bf16 and INT8 ConvRot precision:

📂 ComfyUI/models/
├── 📂 diffusion_models/
│   ├── trellis_2_bf16.safetensors          # or trellis_2_int8_convrot.safetensors
│   └── pixal3d_bf16.safetensors            # or pixal3d_int8_convrot.safetensors
├── 📂 clip_vision/
│   └── dino_v3_vit_l.safetensors           # Pixal3D uses dino_v3_L_naf_fp32.safetensors
└── 📂 vae/
    ├── trellis_2_shape_vae_bf16.safetensors
    └── trellis_2_texture_vae_bf16.safetensors

Comments

Sign in with GitHub to join the discussion.

Loading comments…
TRELLIS.2 and Pixal3D Image-to-3D Generation Now Built into ComfyUI | ComfyUI Wiki