• Your Name's avatar
    Fix diffusers NaN warning by using FP32 instead of FP16 · df8b4875
    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
coderai 299 KB