SonyResearch/micro_diffusion
A research project from Sony that trains 1.16B parameter diffusion models from scratch on a micro-budget ($1,890) using publicly available images.
Not currently ranked — collecting fresh signals.
star history
This repository provides code and methodology for training large-scale diffusion models from scratch on extremely limited budgets. Using approximately 37M real and synthetic images, they train sparse transformer diffusion models achieving FID of 12.7 on zero-shot COCO generation. The codebase includes training scripts, dataset preprocessing utilities, and pre-trained model checkpoints for reproduction.
Frequently asked
- What is SonyResearch/micro_diffusion?
- A research project from Sony that trains 1.16B parameter diffusion models from scratch on a micro-budget ($1,890) using publicly available images.
- Is micro_diffusion open source?
- Yes — SonyResearch/micro_diffusion is open source, released under the Apache-2.0 license.
- What language is micro_diffusion written in?
- SonyResearch/micro_diffusion is primarily written in Python.
- How popular is micro_diffusion?
- SonyResearch/micro_diffusion has 1.6k stars on GitHub.
- Where can I find micro_diffusion?
- SonyResearch/micro_diffusion is on GitHub at https://github.com/SonyResearch/micro_diffusion.