CRAFT: Aligning Diffusion Models with Fine-Tuning Is Easier Than You Think

CRAFT is a new method for improving AI image generation models to better match what humans prefer, using far less data and computing power than existing approaches. Instead of needing thousands of expensive human-labeled examples, CRAFT can work with just 100 samples by first filtering them for quality and then fine-tuning the model, while being 11-220 times faster than competing methods. The researchers also mathematically proved that this simpler approach is actually equivalent to a more complex reinforcement learning method, making it both practical and theoretically sound.