Add noise correction to SDXL

[shauray8]

Addressing singularities present in noise prediction models during the initial sampling step.

---

Paper - None that would describe this other than ImagineFlash
Code - NA

[toandd2]

Hi, may I ask what the motivation behind this work is? It seems amazing, but I don't understand its impact, does it has any research with its..

[shauray8]

Hi @toandd2, I'd suggest reading this https://arxiv.org/pdf/2405.05224 paper to better understand the method, assuming that, from what I think the impact comes from

— is at the initial denoising step if we introduce a bias through the latents, that would directly restrict the latent space (or the direction 🤷‍♂️) the generation can travel through
— and while training, the forward diffusion process makes sure that SNR(T) is 0, so through this we make sure the inference follows that
— this results in more prominent regions through the generations, darker regions are a little more black, more vibrant, also effects like these can be achieved through increasing the guidance scale, but that would result in leaving the training distribution and introduces a lot of artefacts at times IMO

I might be completely wrong through, so a gentle ping to one of the authors from the paper @asanakoy