Practical overview
When it fits
Use MixGRPO when you want research code and a documented experimental workflow for reward-guided FLUX fine-tuning with mixed ODE-SDE sampling.
Who it is for
- Researchers studying reinforcement learning for diffusion and flow-based generative models.
- Practitioners with multi-GPU infrastructure who want to examine the repository's FLUX fine-tuning workflow.
Documented uses
- Explore mixed ODE-SDE sampling in flow-based GRPO experiments.
- Prepare a FLUX.1-dev model, prompt data and reward models for the documented training workflow.
- Configure and invoke the repository's multi-node training script.
Requirements & limits
The repository is publicly visible and identified as open-source. No software license information is supplied.
- The README documents Python 3.12, Conda, an environment setup script and CentOS system packages.
- The workflow calls for FLUX.1-dev, prompt embeddings, HPDv2 prompts and reward-model assets including HPS-v2.1, Pick Score, ImageReward or CLIP Score.
- The specified multi-node training configuration is four nodes with eight GPUs each, with pdsh/torchrun and a Weights & Biases key in the setup.
- The repository links to a MixGRPO checkpoint for the model workflow.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the MixGRPO environment and FLUX model
Set up the documented Conda environment and download FLUX.1-dev for the repository workflow.
Before you start
- Conda
- Python 3.12
- Hugging Face CLI and an account for the FLUX model download
- CentOS package manager access for the listed system packages
Steps
Create and activate the MixGRPO Conda environment with Python 3.12.
Install pdsh, pssh and mesa-libGL using the README's CentOS command, then run bash env_setup.sh.
Create ./data/flux and download black-forest-labs/FLUX.1-dev there using the documented Hugging Face CLI commands.
Expected outcome
The Conda environment and FLUX model files are prepared at the repository's documented path.
Guide limitations
- This guide covers setup only, not training or inference.
- The repository also describes reward-model preparation and prompt preprocessing as later stages.
Configure the documented multi-node training launch
Prepare the repository's host and environment settings before invoking its FLUX GRPO training script.
Before you start
- The MixGRPO environment and model/data prerequisites
- A multi-node setup with host IPs
- A Weights & Biases key
- The README's default resource configuration is four nodes with eight GPUs per node
Steps
Set the multi-node IPs in data/hosts/hostfile.
Run bash scripts/preprocess/set_env_multinode.sh to set INDEX_CUSTOM on each node to 0, 1, 2 and 3, respectively.
Set wandb_key in scripts/finetune/finetune_flux_grpo_FastGRPO.sh.
Invoke bash scripts/finetune/finetune_flux_grpo_FastGRPO.sh.
Expected outcome
The repository's documented training launch is configured and invoked.
Guide limitations
- The README specifies a 32-GPU default configuration; it does not establish that other hardware configurations will work.
- The recipe depends on prepared models, reward models and prompt data.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- Tencent-Hunyuan/MixGRPO · Official homepage title · retained in its source language