Flow-Factory

Why we included this project

Fine-tuning diffusion or flow-matching models usually means stitching together several separate scripts. Flow-Factory folds that into a single config-driven pipeline that handles both online reinforcement learning (GRPO, DPO, DPPO, and newer objectives) and offline paths like SFT and preference-based DPO, with LoRA so you can adapt a model without retraining full weights. The supported models span Stable Diffusion 3.5, FLUX, Qwen-Image, Wan video models, and audio-video generators such as LTX-2 and MiniMax H3, which makes it a practical base for comparing fine-tuning approaches across architectures. A registry-based plugin system lets you add trainers, model adapters, and reward models without touching the core, and the acceleration layer centralizes attention backends, torch.compile, and feature caching. Teams doing research or product work on custom image, video, or audio-video generation will find it a solid starting point rather than a toy demo.