Do the trick related to noise offset

[xiaohu2015]

https://www.crosslabs.org//blog/diffusion-with-offset-noise
@forever208

[forever208]

@xiaohu2015 Thanks for your mentioned interesting blog. DDPM-IP has nothing to do with offset noise.
If you look at their offset noise noise = torch.randn_like(latents) + 0.1 * torch.randn(latents.shape[0], latents.shape[1], 1, 1), they add the same value for each pixel for a channel to implement the drift. Whereas, our is new_noise = noise + 0.1 * th.randn_like(noise).

More importantly, I think their input signal and target signal are both noise = torch.randn_like(latents) + 0.1 * torch.randn(latents.shape[0], latents.shape[1], 1, 1) during training. This is not the case for DDPM-IP