Output has NaN

[ZhongliangGuo]

Hi, I tried the pipe with several images, I found that, when the image has a white background, the pipe may output a black image. There is a warning like:
"diffusers/image_processor.py:97: RuntimeWarning: invalid value encountered in cast images = (images * 255).round().astype("uint8")"

I think that maybe the pipe output a value where not a number? I attached the example image, whatever the random seed I used, it always return a black image. (including my 2 servers, 2 pc, and your hugging face demo)

Could you please help me to solve this problem? I have the same environment as your requirements.txt.

[lightChaserX]

I suppose it is caused by the image with a pure color background and the regularization. You may try to disable the regularization by setting noise_regularization_num_reg_steps to 0

[loggerJK]

You may try this at src/renoise_inversion.py:15

_grad = torch.clip(_grad, -100, 100)
if torch.isnan(_grad).any():
    warnings.warn("NaN detected in l_kld")
else :
    e_t = e_t - lambda_kl * _grad

which does not update e_t if NaN is detected.