Both Trellis.2 (Xiang et al., 2025) and Pixal3D (Li et al., 2026) now run natively in ComfyUI. No custom nodes, no compiled CUDA extensions, no PyTorch downgrades, and no non-commercial dependencies.
This is more than a model integration. It ships with a rebuilt 3D pipeline: new Load/Preview/Save 3D nodes, a set of mesh post-processing nodes, and an extended PBR texturing stage that bakes normal and ambient occlusion maps for a complete material set. Everything runs on consumer hardware, and everything is free to use, including commercially.
Why Trellis.2 still matters, ten months later
When Microsoft open-sourced Trellis.2 in December 2025, it immediately became the best open-source model for 3D generative AI. A 4-billion-parameter model built on a compact structured latent representation (O-Voxel). It generates high-fidelity 3D assets from a single image at effective resolutions up to 1536³, handling complex topologies that earlier methods struggled with. It also shipped with a PBR texturing model generating base color, roughness, and metallic maps.
Ten months is an eternity in generative AI, yet Trellis.2 hasn’t just aged well, it has become foundational. Several open-source 3D models released since build directly on it, the most notable being Pixal3D whose implementation uses the Trellis.2 backbone.
The community got there first
As always, the ComfyUI community was quick to bring Trellis.2 into the graph. Within days of the release, custom node packs appeared, the most popular being ComfyUI-TRELLIS2 by Andrea Pozzetti and ComfyUI-Trellis2 by VisualBruno, which together gathered well over a thousand stars. We’re grateful to both authors as they proved the demand and carried the community for months.
Despite their efforts, running Trellis.2 remained a challenge for two reasons.
Installation
The original implementation targets environments built around PyTorch 2.6.0 with CUDA 12.4, which for many users meant downgrading their existing ComfyUI environment. On top of that sit a stack of compiled CUDA extensions (flash-attention, FlexGEMM sparse convolutions, the O-Voxel kernels, CuMesh, nvdiffrast) each of which must match your exact Python, PyTorch, and CUDA combination. The custom node authors did heroic work shipping prebuilt wheels per configuration, but every PyTorch or CUDA update meant a new round of compilation failures, and installs regularly broke. This is now solved with the native integration in ComfyUI. Follow our installation tutorials for Trellis.2 and Pixal3D.
Licensing
Trellis.2’s own code and weights are MIT-licensed, but its original pipeline depends on NVIDIA’s nvdiffrast (for mesh rasterization) and nvdiffrec (for Physically Based Rendering), both distributed under the NVIDIA Source Code License which restricts usage to non-commercial research and evaluation. In practice, a studio couldn’t ship assets from the reference pipeline without stepping into a legal gray zone. These dependencies have been removed from with the native integration.
Then came Pixal3D
In April 2026, Pixal3D from researchers at Tsinghua University and Tencent ARC Lab got accepted at SIGGRAPH 2026. It pushed open-source 3D generation another step forward with its pixel-aligned generation establishing direct pixel-to-3D correspondences. The result is near-reconstruction-level fidelity to the input view, with detailed geometry and the same PBR material set.
Pixal3D, simplified mesh with PBR material set
Pixal3D is heavily built on Trellis.2 as it uses its backbone and shares its VAEs and DINOv3 image conditioning. This is why integrating it together with Trellis.2 made sense. However Pixal3D generally performs better than Trellis.2 as the generated 3D mesh strictly aligns with the input image.
Model highlights
Trellis.2
Single image to 3D asset. A 4-billion-parameter model that generates high-fidelity geometry and materials from one input image.
O-Voxel structured latents. A native, compact omni-voxel representation encoding both geometry and appearance, generating assets at effective resolutions up to 1536³.
Any topology. Handles open surfaces, non-manifold geometry, and fully-enclosed volumes.
PBR materials built in. A dedicated texturing model generates base color, roughness, and metallic maps.
Pixal3D
Pixel-aligned generation. Geometry is generated in direct correspondence with the input view. What you see in the image is what you get in 3D!
Explicit image back-projection. Multi-scale image features are lifted into a 3D feature volume, delivering near-reconstruction-level fidelity.
Cascaded refinement. A staged process progressively refines sparse structure, shape, and texture up to high resolution.
Built on Trellis.2. Shares the Trellis.2 backbone, VAEs, and DINOv3 conditioning.
Generated results, no post-processing of the mesh, vertex painted
What ships in this integration
The goal was simple: make the best open 3D models run in ComfyUI the way every image or video generation model does. A major thank-you goes to Kijai for the implementation, and to yousef-rafat for the initial draft this work built on. In addition to the native implementation, this has been an opportunity to make 3D generation a first-class citizen in ComfyUI. Here is what shipped:
Pure native implementation
Both Trellis.2 and Pixal3D now run as core ComfyUI nodes. The 3D post-processing that required compiled extensions has been reimplemented from scratch in PyTorch and SciPy. No nvdiffrast, no nvdiffrec, no per-configuration wheels, no PyTorch downgrade. If your ComfyUI runs, these models run on your current PyTorch.
Rebuilt 3D nodes
While these were shipped in an earlier version of ComfyUI, the Load 3D, Preview 3D, and Save 3D nodes have been rebuilt from the ground up to support these models and modern mesh workflows. We’re grateful to Terry Jia for his remarkable work on these nodes. Check out the nodes:
Load 3D (Advanced)
Preview 3D (Advanced)
Save 3D (Advanced)
Native mesh post-processing
Raw generative meshes are rarely production-ready, so this release introduces a new set of post-processing nodes:
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.
Fill Holes: fill-in holes resulting from the generation
And more: Merge Meshes, Paint Mesh, Render Mesh…
A complete PBR texture set
Trellis.2’s texturing model generates base color, roughness, and metallic maps. Our implementation goes further: a new UV unwrapping node prepares the mesh for texturing, and two additional maps are generated: a normal map and an ambient occlusion map, both baked from the high-poly mesh. Are these textures perfect? No. But they’re free, generated on consumer hardware, and yours to use as you wish.
Example results
An honest word on quality
Let’s be direct: the best closed-source 3D generators (Hunyuan 3D, Tripo, Rodin) still produce better results than Trellis.2 and Pixal3D. If you need the highest quality and an API fits your pipeline, those remain strong options (all of them are available through ComfyUI’s partner nodes).
What this integration offers is different: the best open 3D generation available, running locally, at zero cost per asset, with no licensing restrictions on what you make. For iteration, prototyping, stylized work, 3D-to-2D workflows, and anyone who wants full control of their pipeline without spending an afternoon to install.
Getting started
Update ComfyUI to the latest version 0.34.0 or go to Comfy Cloud
Download the workflows below, or find them in the template library.
Follow the note in the workflow to download the models and save them in the correct model directory.
Drop in an image and run.
Model weights:
Comfy-Org/BiRefNet (for background removal)
Comfy-Org/MoGe (for camera FOV estimation)
Спасибо за вашу работу! Как вы и писали выше, это вполне годиться для прототипирования, а я ещё вижу их использование в качестве грубой заготовки для дальнейшей доработки в Синема 4 Д.