One Step Generation
Project details
Description:
Diffusion models have been hugely successful in generating high-quality data, e.g. images. The state-of-the-art pipeline is to first train a diffusion model, and afterwards distil this to a small number - possibly one - step generator. This two-stage training process, whilst pragmatic, raises some deep questions: why can't we train a one-step generator directly, without needing to first train a diffusion model? From the viewpoint of latent variable modelling, we propose to train a one-step generator from scratch, using a more sophisticated posterior variational approximation than a Gaussian, with the view that this may enable us to avoid the diffusion model completely.
Who’s involved?
David Barber, University College London
Sid Narayanaswamy, University of Edinburgh
Expertise sought
This project is seeking collaborators
The project requires experts in probabilistic modelling, with expertise in variational inference and approximate inference. The project also requires coding this (pytorch) to demonstrate results, and eventually to train on larger scale datasets.
Industry is welcome to collaborate. We would value help with coding and running experiments; we would, of course, also welcome input into the methodology and potential application areas beyond image generation.
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
