guided-diffusion VS score_sde

People who made these colabs better and better also the same people who are at Midjourney now. But the "mother" of it all, was Katherine Crowson. She made a fine tuned model that uses a 512x512 unconditional ImageNet diffusion model fine-tuned from OpenAI's 512x512 class-conditional ImageNet diffusion model (https://github.com/openai/guided-diffusion) together with CLIP (https://github.com/openai/CLIP) to connect text prompts with images. It uses a smaller secondary diffusion model trained by Katherine Crowson to remove noise from intermediate timesteps to prepare them for CLIP.

I am trying to use OpenAI's Guided Diffusion Github to train my own diffusion model. I thought to ask here to see if anyone had any experience with it as I've been having trouble training my own models on it. If anyone has any resources to point me towards it would be greatly appreciated!

Our codebase for the diffusion models builds heavily on OpenAI's ADM codebase , lucidrains, Stable Diffusion, Lightning and Hugging Face. Thanks for open-sourcing!

I am working with guided diffusion. I would like to reproduce the results of the repository for the 64->256 super resolution network. https://github.com/openai/guided-diffusion

this code is (mostly) just the original openai guided diffusion code: https://github.com/openai/guided-diffusion

Starting resource, as it is all done through this code (information on how to do it on Colab is out there) https://github.com/openai/guided-diffusion

Original notebook by Katherine Crowson (https://github.com/crowsonkb, https://twitter.com/RiversHaveWings). It uses either OpenAI's 256x256 unconditional ImageNet or Katherine Crowson's fine-tuned 512x512 diffusion model (https://github.com/openai/guided-diffusion), together with CLIP (https://github.com/openai/CLIP) to connect text prompts with images.

Code for https://arxiv.org/abs/2105.05233 found: https://github.com/openai/guided-diffusion

Paper, repo, Colab. It's really good.

Perhaps the ODE interpretation would be helpful (see here and here) which turns DDPMs into neural ODEs using the Fokker-Planck equation so after the initial starting noise, the sampling process is deterministic. If samples are noisy even with the full number of steps then you might need to increase the number of steps further.

Turns out that diffusion models also define a certain differential equation, making it a neural ODE. Then you can just integrate the ODE in the other direction to get the exact inverse for the DDPM (it's not entirely exact b/c of numerical error in the solver, but close enough)

This new approach to generative modelling looks very intriguing.

In a similar ilk, there's this ICLR paper from this year using stochastic differential equations for generative modelling: https://arxiv.org/abs/2011.13456