Stop rewriting RL glue for every new diffusion model

What it does Flow-Factory is a training framework that wraps popular diffusion and flow-matching generators—FLUX, Stable Diffusion 3.5, Qwen-Image, Wan video models, and even LTX-2 audio-video—under one reinforcement-learning interface. You pick a model, pick an algorithm like GRPO or DPO, and the framework claims every combination works out of the box. It handles the dataset formatting, reward plumbing, and experiment tracking so you can focus on the actual policy optimization.

The interesting bit The real win is the decoupling: models and algorithms are fully independent, meaning you don’t need a custom training script for each new architecture or RL method. The framework also abstracts away attention-backend swaps and bundles unreleased diffusers patches for bleeding-edge models like LTX-2, which suggests the maintainers are eating their own dogfood on the frontier.

Key highlights

• Supports eight RL algorithms including GRPO, DPO, DiffusionNFT, AWM, DGPO, and CRD, with model–algorithm pairs verified in example configs.
• Covers text-to-image, image-to-image, text-to-video, image-to-video, and synchronized audio-video generation across more than a dozen model families.
• Ships with a bundled diffusers submodule for models whose pipeline code hasn’t landed in the official release yet.
• Logs training samples and metrics to Weights & Biases, SwanLab, or TensorBoard without extra instrumentation.
• Provides guidance docs for adding new models and custom reward functions, plus a Docker image for CUDA training environments.

Caveats

• The attention-backend switching is marked experimental and leans on diffusers internals that may shift.
• Some models require the bundled diffusers submodule because upstream support is still pending.

Verdict Researchers and engineers who regularly fine-tune generative models with RL feedback should keep this on hand; if you only run occasional LoRA jobs with static datasets, it is probably overkill.