Finetune Stable Diffusion Models with DDPO via TRL
Hugging Face Blog
Hugging Face's TRL library integrated DDPO (Denoising Diffusion Policy Optimization), a reinforcement learning method for fine-tuning diffusion models like Stable Diffusion to align outputs with human preferences. The method requires an A100 GPU minimum and uses a reward model trained on aesthetic preferences, with recommended hyperparameters including a learning rate of 3e-4 and training batch size of 3 for single-GPU setups. Users can now fine-tune Stable Diffusion models to generate images matching specific objectives such as visual aesthetics without needing the supervised fine-tuning step typically required in RLHF workflows.
