Get the latest tech news
Diffusion training from scratch on a micro-budget
Official repository for our work on micro-budget training of large-scale diffusion models. - SonyResearch/micro_diffusion
This repository provides a minimalistic implementation of our approach to training large-scale diffusion models from scratch on an extremely low budget. In particular, using only 37M publicly available real and synthetic images, we train a 1.16 billion parameter sparse transformer with a total cost of only $1,890 and achieve an FID of 12.7 in zero-shot generation on the COCO dataset. Using the patch-mixer significantly reduces performance degradation with masking while providing a large reduction in training time.
Or read this on Hacker News
