Single GPU inference performance - plans for quantized/optimized version?

[aosteklov]

Hi, first of all great work on Vista4D — really impressive results!

I've been trying to run inference on a single RTX 5090 (32GB VRAM) and
finding it extremely slow due to the 14B Wan model constantly offloading
between VRAM and system RAM (the full model is ~64GB in bfloat16).

A few questions:

1. Are there plans to release an fp8 or quantized version of the checkpoint
   that would fit within 32GB VRAM?
2. Is there a recommended minimum VRAM for reasonable single-GPU inference
   speeds?
3. Would a ComfyUI integration be something you're considering?

Happy to test any experimental builds. Thanks!