Fix: Pixal3D Windows compatibility + configurable NAF target size

[acolasialiuliu]

Summary

This PR introduces Windows compatibility fixes for Pixal3D (TencentARC/Pixal3D) rendering on ComfyUI, and adds a configurable NAF target size parameter.

Changes

1. CPU/CUDA Device Mismatch Fix

File: trellis2/pipelines/trellis2_image_to_3d.py

• get_proj_cond_ss() and get_proj_cond_shape() now check if the image_cond_model is on CPU before moving to CUDA, and only move it back to CPU in low_vram mode if it was originally on CPU. This prevents RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same when the model is loaded without low_vram mode.

2. NAF Model Graceful Fallback

File: trellis2/trainers/flow_matching/mixins/image_conditioned_proj.py

• Added self.naf_download_if_missing = True to DinoV3ProjFeatureExtractor.__init__()
• Wrapped NAF model loading in try/except — on failure, sets self.naf_model = False instead of crashing with AttributeError
• Updated forward(), to(), and cuda() methods to handle naf_model being False (not just None)

This allows Pixal3D to run even when NATTEN is unavailable (NAF will be disabled automatically).

3. Configurable NAF Target Size

File: nodes.py

• Added "naf_target_size" widget to Trellis2ImageCondGenerator.INPUT_TYPES() (INT, default 256, range 128-1024)
• Pipeline's __init__ now has self.naf_target_size_override = None
• Added _get_pixal3d_config() helper method that copies stage config with override applied
• Lower NAF target size significantly reduces VRAM usage on consumer GPUs (16GB)

4. NATTEN Windows Resources & Contact

• Added reference comment in nodes.py:
  • NATTEN Windows wheels: https://hf-mirror.com/lldacing/NATTEN-windows/tree/main
  • Contact: https://space.bilibili.com/37411464

Tested Environment

• Windows 64-bit, Python 3.12.10, PyTorch 2.7.0+cu128
• NVIDIA RTX 4070 Ti SUPER (16GB VRAM)
• NATTEN 0.17.5 (community Windows wheel)
• Flash Attention 2.8.3