# yifan123/flow_grpo

[NeurIPS 2025] An official implementation of Flow-GRPO: Training Flow Matching Models via Online RL

Repository: https://github.com/yifan123/flow_grpo
Canonical: https://ross.abutalabs.com/products/flow_grpo
Homepage: https://arxiv.org/pdf/2505.05470
Language: Python
License: MIT
License Family: permissive
Last push: 2026-05-07T08:23:21+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 81, release rhythm 35, longevity 34
- inputs: {"age_days": 482, "days_push": 118, "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 2498, forks 167 (observed 2026-08-28T04:06:56.795720+00:00)

## What it is
Flow-GRPO is the official PyTorch implementation of a NeurIPS 2025 paper that trains flow matching models (e.g., SD3.5, FLUX.1, Qwen-Image, Wan2.1) via online reinforcement learning with GRPO. It provides training scripts, reward model integrations (CLIPScore, Geneval, OCR), and a faster Flow-GRPO-Fast variant.

## Use cases
- fine-tune a flow matching diffusion model with online RL
- train Stable Diffusion 3.5 with GRPO rewards
- improve text-to-image prompt adherence using CLIPScore reward
- train FLUX.1-dev with reinforcement learning
- apply RL post-training to Qwen-Image or Wan2.1 video models
- train image editing models with Geneval counting rewards
- speed up RL fine-tuning with few-step Flow-GRPO-Fast

## When to choose
- you want to RL fine-tune a flow matching generative model with custom rewards
- you need a research-grade reference implementation of GRPO for diffusion models
- you work with supported models like SD3.5, FLUX, Qwen-Image, or Wan2.1 and have GPU resources

## When to avoid
- you need a production training framework with broad model support out of the box
- your models are not flow matching based (e.g., classic DDPM diffusion)
- you lack multi-GPU hardware or don't want to write custom reward functions

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-training, reinforcement-learning, stable-diffusion, gpu-computing
- domain: machine-learning, deep-learning, image-processing, artificial-intelligence
- platform: python
- tags: flow-matching, grpo, online-rl, diffusion-models, text-to-image, research-code, neurips-2025, gpu, linux

## Member repositories
- yifan123/flow_grpo (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:56.795720+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 2026-08-30T02:27:03.868132+00:00, confidence not recorded.
  - readme: https://github.com/yifan123/flow_grpo (fetched 2026-08-28T04:06:56.795720+00:00, sha f5bd7c372ed4)
- Data as of 2026-08-30T08:39:29.467469+00:00.