Single-image 3D generation has a characteristic failure: the front looks like your reference and everything else is a guess. Textures warp around the sides, the back is invented, and the silhouette drifts from what you asked for.
Pixal3D, from Tencent ARC and presented at SIGGRAPH 2026, attacks the first half of that problem with a specific trick, and its new multi-view mode attacks the second.
Pixel-aligned means what it says
Pixal3D establishes direct pixel-to-3D correspondence through back-projection. The consequence is concrete: the front of the generated model matches your input image 1:1 — no warped or misaligned textures on the face you actually care about.
It outputs a high-fidelity 3D asset with full PBR materials, not just a coloured mesh, which is the difference between an asset you can light in a scene and a novelty you screenshot.
The multi-view part is the news
One image can’t tell you what the back looks like. So:
- September 1, 2026 — Pixal3D released its multi-view inference code
- September 8, 2026 — ComfyUI core landed native support for the multi-view weights (PR #16048)
You can now feed up to four orbit shots of the same object, and the new Pixal3DMultiViewConditioning node combines them into a single, more accurate mesh.
Four orthographic views is, not coincidentally, how character and prop turnarounds have been drawn for decades. The workflow now matches how reference material already exists.
It’s genuinely in core
This matters more than it sounds. Pixal3D’s ComfyUI integration merged into core on August 22, 2026 — so:
- No custom node pack
- No compiled CUDA extensions
- No PyTorch downgrades
- No non-commercial dependencies
Anyone who has spent an evening resolving a custom-node dependency tree to try a 3D model will understand why that list is the selling point. It runs on consumer hardware; community GGUF builds have it working at 6GB VRAM.
The tracker signal, and a licence caveat
Comfy-Org/Pixal3D entered our tracked set this week at 612,886 downloads — the second-largest new entrant on today’s board. That’s the adoption curve catching up to a release that happened three weeks earlier, which is the usual shape: core integration lands quietly, then the tutorials appear, then the downloads.
On licensing, be careful and check for yourself. The Comfy-Org repackaged weights are published under MIT, and ComfyUI’s own announcement stresses no non-commercial dependencies. But at least one third-party ComfyUI node repo for Pixal3D describes the upstream model licence as academic-only. Those two claims cannot both be right about the same artifact, and we could not resolve which applies to the upstream Tencent ARC release versus the Comfy-Org repackage.
If your use is commercial, read the licence on the exact weights you download rather than trusting either a blog post or this article. This is the second time this week — after Qwen-Image-2.1 — that the licence on a repackaged model has been the least clear thing about it.
Related Reading
- Pixal3D Multiview in ComfyUI | Load & Save 3D Nodes — pixaroma (YouTube)
- Comfy-Org/Pixal3D — Hugging Face
- Trellis.2 and Pixal3D Are Now Native in ComfyUI — ComfyUI blog
- Pixal3D Multi-View Comes to ComfyUI: Better 3D From 4 Views — ComfyUI Wiki
- Pixal3D Image to 3D Model ComfyUI Workflow Example — ComfyUI docs
- Create Accurate 3D Characters from Four Views Locally in ComfyUI (YouTube)