Divide-and-Denoise: A Game-Theoretic Method for Fairly Composing Diffusion Models

The article introduces the Divide-and-Denoise method, which coordinates multiple pre-trained diffusion models during sampling by creating a fair division of labor among them. This approach involves solving a fair division game to allocate responsibilities for denoising across different regions of a noisy sample, enhancing the overall utility while maintaining fairness. Evaluations on conditional image generation demonstrate that Divide-and-Denoise outperforms existing methods on quality metrics, effectively leveraging the strengths of each model and addressing common issues like missing objects and mismatched attributes, making it a valuable technique for practitioners working with composite diffusion models.