TencentARC/Pixal3D

Thanks for raising this. For Windows users, there is already a community implementation that provides Windows-related installation instructions which may be useful as a reference:

https://github.com/Saganaki22/Pixal3D-ComfyUI

In particular, the README includes a “Windows Wheel Order” section covering PyTorch CUDA, the required Pixal3D CUDA wheels (`flex_gemm_ap`, `cumesh_vb`, `o_voxel_vb_ap`, `drtk`), FlashAttention, and optional NATTEN/libnatten setup. It also includes recommended Windows loader settings and fallback options for NATTEN/NAF.

This is a ComfyUI integration rather than an official setup guide, but it may help Windows users get Pixal3D running while official `uv` / Windows environment support is still being considered.

> After several tests on my end, I'm also reaching the same conclusion. I tested it on a chair, and the chair legs have incorrect dimensions. Trellis in normal mode gives better results — the solution would be multi-image, but we'd need to manage to get a good calibration. I also tried using [anycalib](https://huggingface.co/spaces/simonpick/anycalib-gpu) to remove the distortion from the images and improve the result, but it doesn't change much.

Thanks for the testing. Could you please provide the original chair input image? We can take a closer look at the issue.

> Input：
>
>
> output:
>
>

there is no hope to run it on windows as you use NATTEN and they refuse to support windows even when local compiled...
and no there are no wheels compatible with the current trellis torch version

After several tests on my end, I'm also reaching the same conclusion. I tested it on a chair, and the chair legs have incorrect dimensions. Trellis in normal mode gives better results — the solution would be multi-image, but we'd need to manage to get a good calibration. I also tried using [anycalib](https://huggingface.co/spaces/simonpick/anycalib-gpu) to remove the distortion from the images and improve the result, but it doesn't change much.

Thanks for the feedback!

For the environment setup, we have been using the current TRELLIS.2-based configuration internally and it works on our side. However, we have not tested the setup on Windows, since we currently do not have a corresponding Windows environment for validation. If you have a working `uv`-based setup or Windows-specific fixes, PRs are very welcome!

Regarding the HF demo, we have deployed the Space using our own GPU resources. You can click the **“Free Deploy Server”** option on the left side of the Space to run the demo without using your own credits.

Hope this helps, and enjoy!

Thanks ，but your a little advice, wheels are better to be in a new project, too big. Too slow to pull in china, github is almost banned in china. And most of the wheels are useless for one platform

You can try it using ComfyUI-Trellis2, branch pixal3d: https://github.com/visualbruno/ComfyUI-Trellis2/tree/pixal3d
It's in beta, I'm still working on it, but it works

感谢您的关注和支持！Pixal3D在一致性上确实有一定的优势，但在细节质量、表面光滑度、结构合理性等维度均明显弱于Hunyuan3D 3.1等闭源模型。如果Pixal3D能带来一些新的想法，或在特定场景下有更适合的应用，就已经很好了。

我们会去调研一下comfyui的实现可能性，感谢您的建议！