-
Your Name authored
- Changed torch_dtype from float16 (when CUDA available) to float32 - This prevents NaN/Infinity values in image output that cause black/corrupted images - FP16 can cause numerical overflow on some models like SDXL
df8b4875
| Name |
Last commit
|
Last update |
|---|---|---|
| .vscode | ||
| .gitignore | ||
| LICENSE.md | ||
| README.md | ||
| aaa | ||
| build.sh | ||
| coder | ||
| coderai | ||
| requirements-nvidia.txt | ||
| requirements-vulkan.txt | ||
| requirements.txt | ||
| requirements.txt~ |