Fix inconsistency in CUDA rasterizer option causing startup failures in WSL

[xhiroga]

Problem Description

When running in WSL environment, ComfyUI-3D-Pack fails to load with the following error:

RuntimeError: Cuda error: 304[cudaGraphicsGLRegisterBuffer(&s.cudaPosBuffer, s.glPosBuffer, cudaGraphicsRegisterFlagsWriteDiscard);]

Root Cause

The issue stems from an inconsistency in how the rasterizer is selected:

1. The multiview_color_projection_texture() function offers a force_cuda_rast parameter that allows choosing between OpenGL and CUDA rasterizers
2. However, the Pix2FacesRenderer class in project_mesh.py always initializes with RasterizeGLContext during module loading
3. This initial OpenGL-based renderer creation fails in WSL2 due to OpenGL-CUDA interoperability limitations (like Unable to load under WSL 2 in Win10 due to: RuntimeError: Cuda error: 304[cudaGraphicsGLRegisterBuffer(&s.cudaPosBuffer, s.glPosBuffer, cudaGraphicsRegisterFlagsWriteDiscard);] #425)

Proposed Fix

The solution is to make the CUDA rasterizer option consistent throughout the codebase:

class Pix2FacesRenderer:
    def __init__(self, device="cuda"):
      try:
        self._glctx = dr.RasterizeGLContext(output_db=False, device=device)
      except Exception:
        self._glctx = dr.RasterizeCudaContext()

      self.device = device
      _warmup(self._glctx, device)

I've verified that this approach works by manually modifying the code. However, I would appreciate any feedback on this fix before submitting a PR.
I would also be happy to submit a PR with this fix if it looks good.

[rob-ack]

I want to ref that this is also able to fix #425 (which is closed as not supported). I would like to see support for WSL so good job so far 👍

[HDANILO]

I support the fix, it's rather annoying error to be manually fixing everytime

[Ainaemaet]

Has this fix been committed? I haven't tried the nodes yet out of fear of issues, but would like to.

[DenisKochetov]

thank you guys it was fixed
a9d4562