stable-dreamfusion
stable-dreamfusion copied to clipboard
SDS loss
According to section 2 in the paper, epshat = pred_noise_text + guidance_scale * (pred_noise_text - pred_noise_uncond)
.
While it seems that in the implementation in sd.py, it's assigned as: epshat = pred_noise_uncond + guidance_scale * (pred_noise_text - pred_noise_uncond)
.
Is this intentional?
@roibaron Hi, you may check this. Since the guidance_scale is set to a very large value, this change of order may not have a siginificant influence.