-
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
- 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