# sihyun-yu/REPA

[ICLR'25 Oral] Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think

Repository: https://github.com/sihyun-yu/REPA
Canonical: https://ross.abutalabs.com/products/repa
Homepage: http://sihyun.me/REPA/
Language: Python
License: MIT
License Family: permissive
Topics: diffusion-models, diffusion
Last push: 2025-03-16T13:50:00+00:00

## Health v2 (maintenance only)
Score: 27/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 11, release rhythm 35, longevity 49
- inputs: {"age_days": 695, "days_push": 535, "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 1700, forks 98 (observed 2026-08-28T04:05:24.215046+00:00)

## What it is
REPA is the official PyTorch implementation of the ICLR 2025 paper 'Representation Alignment for Generation', a regularization technique that aligns diffusion transformer representations with pretrained self-supervised visual encoders. It dramatically accelerates training of diffusion models like SiT (17.5x faster convergence) and achieves state-of-the-art image generation quality (FID=1.42 on ImageNet 256x256).

## Use cases
- train diffusion transformers faster with representation alignment
- reproduce REPA results on ImageNet 256x256
- improve generation quality of SiT or DiT models
- apply representation alignment regularization to diffusion model training
- research denoising-based generative models
- benchmark image generation FID scores

## When to choose
- you are training diffusion transformers or flow-based models and want faster convergence
- you need state-of-the-art class-conditional ImageNet generation results
- you are doing research on representation learning in diffusion models

## When to avoid
- you need a production-ready image generation service rather than research training code
- you want text-to-image generation out of the box (experiments are ImageNet class-conditional)
- you need a general-purpose diffusion library like Diffusers

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, image-processing
- domain: deep-learning, machine-learning, image-processing
- platform: python
- tags: diffusion-models, diffusion-transformers, representation-alignment, research-code, image-generation, iclr-2025, training-efficiency, gpu, linux

## Member repositories
- sihyun-yu/REPA (main) score 27

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:24.215046+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 2026-08-30T03:37:46.774618+00:00, confidence not recorded.
  - readme: https://github.com/sihyun-yu/REPA (fetched 2026-08-28T04:05:24.215046+00:00, sha 641175056a2c)
  - homepage: http://sihyun.me/REPA/ (fetched 2026-08-29T11:12:25.444698+00:00, sha 5e447b8e9c30)
- Data as of 2026-08-30T08:39:29.467469+00:00.