Divide-and-Denoise: A Game-Theoretic Method for Fairly Composing Diffusion Models
Researchers have developed a new method called Divide-and-Denoise for combining multiple pre-trained diffusion models. This technique uses game theory to fairly divide the denoising task among different models, preventing any single model from dominating and ensuring better cooperation. The method has been evaluated on conditional image generation tasks and shows improved performance over existing approaches, addressing issues like missing objects and attribute mismatches. AI
IMPACT This method could lead to more robust and diverse image generation by effectively leveraging multiple specialized diffusion models.