Classifier-free guidance

[ThibaultGROUEIX]

Hi,

Just going over the code line-by-line and checking things ;) this line

I think

noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

should be

noise_pred = noise_pred_text + guidance_scale * (noise_pred_text - noise_pred_uncond)

to match the equation of the paper

Happy to do a PR if you confirm it needs to be change.

Thanks,

[ashawkey]

@ThibaultGROUEIX Hi, thanks for the checking! I simply followed diffusers' tutorial here. You can find it in the last part.

[ThibaultGROUEIX]

indeed ! Seems like either the paper or the tuto has a typo. I a not sure which. It may not matter in this case since the guidance parameter is so high.

[roibaron]

@ThibaultGROUEIX Did you make this change work?

[chenguolin]

There are two forms to express classifier-free guidance:

1. epsilon = epsilon_uncond + s * (epsilon - epsilon_uncond), which is used in GLIDE (Nichol et al., 2022), LDM and Stable Diffusion (Rombach et al., 2022)
2. epsilon = (1+w) epsilon + w epsilon_uncond, which is used in the classifier-free guidance original paper (Ho and Salimans, 2021) and DreamFusion (Poole et al., 2022)

Both of them are correct. But for the first case, you should set s>1 to enable classifier-free guidance, and set w>0 instead in the second case.