# SonyResearch/micro_diffusion

Official repository for our work on micro-budget training of large-scale diffusion models.

Repository: https://github.com/SonyResearch/micro_diffusion
Canonical: https://ross.abutalabs.com/products/micro_diffusion
Homepage: https://arxiv.org/abs/2407.15811
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-01-12T20:19:19+00:00

## Health v2 (maintenance only)
Score: 24/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 1, release rhythm 35, longevity 55
- inputs: {"age_days": 778, "days_push": 598, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1594, forks 57 (observed 2026-08-28T04:05:08.982785+00:00)

## What it is
Official implementation of Sony Research's micro-budget approach to training large-scale text-to-image diffusion transformer models from scratch. It trains a 1.16B parameter sparse transformer for about $1,890 using patch masking, mixture-of-experts layers, and a mix of real and synthetic images.

## Use cases
- train a text-to-image diffusion model on a small budget
- reproduce low-cost diffusion transformer training from a paper
- generate images with a pretrained micro-budget DiT checkpoint
- experiment with patch masking for diffusion training
- train diffusion models on a small image dataset subset
- fine-tune a diffusion model from 256x256 to 512x512 resolution

## When to choose
- you want to train or fine-tune a text-to-image diffusion model with limited GPU budget
- you need reproducible research code for micro-budget diffusion training
- you want to study patch masking and mixture-of-experts in DiT models

## When to avoid
- you need a production-ready image generation API or polished inference tool
- you lack multi-GPU hardware (reference setup uses 8x H100)
- you want general-purpose diffusion fine-tuning of existing models like Stable Diffusion rather than training from scratch

## Facets
- artifact type: library
- maturity: active
- function: deep-learning, machine-learning, image-processing, llm-training
- domain: deep-learning, machine-learning, image-processing, artificial-intelligence
- platform: python
- tags: diffusion-models, text-to-image, micro-budget-training, patch-masking, mixture-of-experts, dit, research-code, image-generation, gpu, linux

## Member repositories
- SonyResearch/micro_diffusion (main) score 24

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:08.982785+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:54:10.700649+00:00, confidence not recorded.
  - readme: https://github.com/SonyResearch/micro_diffusion (fetched 2026-08-28T04:05:08.982785+00:00, sha 703698602269)
  - homepage: https://arxiv.org/abs/2407.15811 (fetched 2026-08-29T11:24:59.547534+00:00, sha 4c9192299e9e)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T11:24:59.557053+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T11:24:59.560602+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T11:24:59.562408+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T11:24:59.558880+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.