# XueZeyue/DanceGRPO

An official implementation of DanceGRPO: Unleashing GRPO on Visual Generation

Repository: https://github.com/XueZeyue/DanceGRPO
Canonical: https://ross.abutalabs.com/products/dancegrpo
Homepage: https://arxiv.org/abs/2505.07818
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-10-16T12:56:47+00:00

## Health v2 (maintenance only)
Score: 40/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 47, release rhythm 35, longevity 34
- inputs: {"age_days": 482, "days_push": 321, "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 1648, forks 82 (observed 2026-08-28T04:05:16.646919+00:00)

## What it is
Official implementation of DanceGRPO, a framework applying Group Relative Policy Optimization (GRPO) to fine-tune visual generation models with reinforcement learning. Built on FastVideo, it supports diffusion and rectified-flow models including Stable Diffusion, FLUX, HunyuanVideo, SkyReels-I2V, Qwen-Image, and Wan-2.1.

## Use cases
- fine-tune FLUX with human preference reward models
- apply GRPO reinforcement learning to diffusion models
- align text-to-image models with human preferences
- train video generation models with RL
- optimize image aesthetics with reward models
- reproduce the DanceGRPO paper experiments

## When to choose
- you need RL-based preference alignment for diffusion or rectified-flow image/video models
- you want to fine-tune FLUX, Stable Diffusion, HunyuanVideo, or Wan-2.1 with GRPO
- you are researching RL fine-tuning of generative visual models

## When to avoid
- you need general-purpose LLM RLHF rather than visual generation
- you lack GPU resources for large model training
- you want a plug-and-play inference tool rather than training scripts

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, deep-learning, image-processing, video-processing
- domain: machine-learning, deep-learning, image-processing, artificial-intelligence
- platform: python
- tags: grpo, reinforcement-learning, diffusion-models, rectified-flow, text-to-image, text-to-video, reward-models, fine-tuning, stable-diffusion, flux, hunyuanvideo, research-code, video, gpu, linux

## Member repositories
- XueZeyue/DanceGRPO (main) score 40

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:16.646919+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 2026-08-30T03:45:12.090024+00:00, confidence not recorded.
  - readme: https://github.com/XueZeyue/DanceGRPO (fetched 2026-08-28T04:05:16.646919+00:00, sha ee459f39c887)
  - homepage: https://arxiv.org/abs/2505.07818 (fetched 2026-08-29T11:18:28.658090+00:00, sha f3985970efa3)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T11:18:28.667303+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T11:18:28.670806+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T11:18:28.672763+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T11:18:28.669186+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.