Diff-Instruct++: Training One-step Text-to-image Generator Model to Align with Human Preferences [Huggingface][Colab]

Welcome to the official repository for Diff-Instruct++ (DI++), a novel preference alignment approach for DiT-based 1-step text-to-image generative models. Diff-Instruct* is built upon a new KL-based RLHF theory, which improves human preferences while maintaining the advantage of 1-step generation.