Pixal3D ComfyUI Guide

What "Pixal3D ComfyUI" refers to

Pixal3D is TencentARC's open-weights image-to-3D model: one image goes in, a textured GLB with PBR materials comes out. It ships as a Python project, not a ComfyUI pack, so everything under the "Pixal3D ComfyUI" search term is a community wrapper that loads those same weights behind ComfyUI nodes.

Three wrappers are in circulation as of November 2026:

They all wrap the same upstream weights, so the reconstruction you get out of any of them is the same pipeline described on 3D Pixel and open-source image to 3D model. What differs is the install pain and the memory knobs.

Check your hardware before you clone anything

The wrappers need CUDA. AMD and Intel GPUs are not supported by any of the three today, and Apple Silicon has no working path — the compiled extension wheels are CUDA-only.

If your card is not on this list, the honest answer is that local ComfyUI is the wrong lane and a hosted run is cheaper than the upgrade.

VRAM and wall-clock per quality tier

These figures come from the dreamrec/ComfyUI-Pixal3D README rather than from us — treat them as the wrapper author's measurements on an RTX 5090, not as a guarantee for your card:

Two costs sit outside that table. The pipeline occupies roughly 14 GB just sitting loaded if you keep the model warm, and a cold start pays one to three minutes of weight loading before the first render. A use_tiled_decoder switch can lower the peak enough to reach 1536 on a 24 GB card.

Install order that actually works

1. Confirm the ComfyUI build is current and that torch reports a CUDA version matching the wheels you intend to install. A mismatch here is the single most common failure.
2. Install the wrapper through ComfyUI Manager, or clone it into ComfyUI/custom_nodes/.
3. Install the few extra GPU extensions the wrapper lists — this is where installs stall. flex_gemm, o_voxel and nvdiffrast all compile against your CUDA toolkit; there is no pure-Python fallback.
4. Install the matching NATTEN wheel, then restart ComfyUI and check its startup log for import errors before loading a workflow.
5. Download the weights into the folder the node pack names, then run one 1024_cascade job at the lowest token budget to prove the graph before you attempt 1536.

Failures worth recognizing on sight

• ImportError on flex_gemm or o_voxel — the extension did not compile. It is a toolchain problem, not a model problem; reinstalling the node pack will not fix it.
• OOM at 1536_cascade with low_vram=false — expected on 24 GB cards. Either turn low_vram on or stay at 1024.
• Sharp source image, soft back faces — this is the model's boundary, not a setting. Pixel-aligned reconstruction cannot invent pixels your photo never contained; the photo to 3D model guide covers how to frame around it.
• A node that imports but produces an untextured mesh — usually the texture decoder was skipped when the wrapper's submodules were not fetched with the clone.

When the local path stops being the cheap one

A working 48 GB box, the hours spent on CUDA extensions, and the 14 GB of resident memory are the real price of the local lane. If you need one model a week rather than a hundred, the hosted workspace on this site runs the same pixel-aligned pipeline and returns the same kind of GLB: open the AI 3D generator, or start from the free image to 3D model page to see what one photo produces before you pay for anything.

Whichever lane you pick, the input rules are identical — the 2D image to 3D model and image to 3D mesh guides list them, and 3D AI explains where Pixal3D stops and a generative model like Trellis 2 begins.

FAQ

Is there an official Pixal3D ComfyUI node?

No. TencentARC publishes the model and inference code; every ComfyUI integration is maintained by a third party. Pin the wrapper version you install, because the three known packs update their wheel requirements independently.

Can I run Pixal3D in ComfyUI without a GPU?

Not in any of the current wrappers. They compile CUDA extensions at install time. Without an NVIDIA card, a hosted or browser-based run is the only path.

Does the ComfyUI version produce a different model than the hosted tool?

It runs the same upstream weights, so geometry and texture behaviour match. Practical differences come from the settings: the wrappers expose token budgets and cascade resolution that change detail level, which is exactly why the VRAM table above exists.

16 GB enough?

For a 1024_cascade run with reduced tokens, yes — the wrapper author measured about 10 GB at 8/8/8 and about 17 GB at 16/16/16. Expect minutes per image rather than seconds, and do not plan on 1536_cascade.