ReNFT Repairs Mode Collapse in Diffusion Models, Enhancing Diversity.

Diffusion models, when fine-tuned with rewards, often suffer from "mode collapse," where they generate highly similar outputs for a given prompt, losing diversity. Existing solutions typically involve adding external signals or modifying the model's input. This new research introduces ReNFT, a technique that addresses mode collapse from within the diffusion generator itself. ReNFT operates on the principle that mode collapse isn't about deleting diverse capabilities but rather suppressing them. It uses unconditional probes to identify "anti-hub" prompts where the model's inherent biases are most apparent. Then, it generates counterfactual proposals and uses reward ranking to assign "pull" and "push" roles, effectively recalibrating the internal probability distribution. Empirical results show that ReNFT maintains nearly all of the reward performance of previous methods while dramatically improving diversity metrics. This internal repair mechanism offers a novel and complementary approach to enhancing the robustness and creative range of diffusion models.