OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Developer Tools & SDKs

Unirl

Unirl is a Python reinforcement-learning framework for jointly training language models and flow-matching image models, with language models refining prompts during text-to-image generation or image-editing training.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Unirl — OpenViST project profile in AI Developer Tools & SDKs.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Unirl when you need a research framework for reinforcement-learning experiments that train prompt refinement and flow-matching image generation or editing together.

Who it is for

  • Researchers experimenting with reinforcement-learning methods for text-to-image generation.
  • Developers training prompt-refining language models alongside FLUX image models.
  • Teams exploring reinforcement-learning workflows for image editing with EditReward.

Documented uses

  • Train a language model to adaptively refine prompts within an image-generation reinforcement-learning loop.
  • Run prompt-only training that updates the language model while keeping the diffusion transformer frozen.
  • Run joint training that updates both the language model and image model.
  • Experiment with image-editing training using EditReward and supported FLUX backbones.
Before you choose

Requirements & limits

The repository is public and identifies its license as Apache-2.0. The documented training setup requires Linux, NVIDIA GPUs, Python 3.11, PyTorch 2.7 and CUDA 12.x. Some model and reward checkpoints are gated on Hugging Face and require license acceptance and login before download.

  • The README describes both prompt-only training, with the image transformer frozen, and joint language-model plus image-model training.
  • The default training and reward configuration uses eight GPUs total.
  • The image-editing path requires the separately run EditReward service and source-image/edit-prompt training data.
  • The README’s qualitative evaluation command runs evaluation jobs and is explicitly not an installation smoke test.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install the UniRL training environment

Create the documented training environment on Linux for UniRL’s reinforcement-learning workflows.

Before you start

  • Linux with NVIDIA GPUs
  • Python 3.11
  • PyTorch 2.7 with CUDA 12.x
  • Conda

Steps

  1. Create the environment with conda env create -f environment.yml.

  2. Activate it with conda activate unirl.

  3. Install OpenAI CLIP with the command below.

    Documented command
    pip install git+https://github.com/openai/CLIP.git

  4. Install diffusers>=0.38.0 and huggingface-hub==0.36.2, then install flash-attn==2.7.4.post1 using MAX_JOBS=4 and --no-build-isolation.

  5. Before downloading gated FLUX or reward checkpoints, authenticate with huggingface-cli login.

Expected outcome

The documented UniRL training environment and required packages are installed; Hub authentication is available for gated checkpoint downloads.

Guide limitations

  • This is an environment setup guide, not an installation smoke test or training run.
  • The documented default training and reward configuration uses eight GPUs total.
  • The listed FLUX and reward checkpoints are gated on the Hugging Face Hub.
Documentation 1 ↗Documentation 2 ↗
Source-based practical guide

Run the documented GenEval prompt-only training launch

Launch prompt-only GenEval training with the QwenFlux checkpoint and FLUX.1-dev.

Before you start

  • A Linux NVIDIA-GPU machine with the documented UniRL training environment installed.
  • A built QwenFlux checkpoint at outputs/pretrain/qwenflux.
  • A running GenEval reward service reachable at http://127.0.0.1:18085/.
  • The repository checked out locally, with its documented scripts available.

Steps

  1. Set MODEL_NAME_OR_PATH=outputs/pretrain/qwenflux.

  2. Set GENEVAL_URL=http://127.0.0.1:18085/ and RUN_NAME=qwenflux-geneval-prompt.

  3. Launch the documented script with the command below.

    Documented command
    bash scripts/train/geneval_grpo_flux.sh

Expected outcome

The documented prompt-only GenEval training launch is configured to update the language model while keeping the DiT frozen.

Guide limitations

  • This recipe covers the prompt-only launch, not installation of the GenEval reward service or building the checkpoint.
  • The source describes a separate joint-training variant that sets JOINT=true; this recipe does not launch that variant.
  • No successful training run is established by the documentation.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
G-U-N/UniRL · Official homepage title · retained in its source language
How we review sourcesSuggest a correction