Censored Sampling of Diffusion Models Using 3 Minutes of Human Feedback
This addresses the issue of partial misalignment in diffusion models for image generation, offering a practical solution for users needing to filter out bad outputs with minimal human input.
The paper tackles the problem of preventing undesirable image generation from pre-trained diffusion models, known as censoring, by using a reward model trained on minimal human feedback. They demonstrate that censoring can be achieved with extreme efficiency, requiring only a few minutes of human feedback.
Diffusion models have recently shown remarkable success in high-quality image generation. Sometimes, however, a pre-trained diffusion model exhibits partial misalignment in the sense that the model can generate good images, but it sometimes outputs undesirable images. If so, we simply need to prevent the generation of the bad images, and we call this task censoring. In this work, we present censored generation with a pre-trained diffusion model using a reward model trained on minimal human feedback. We show that censoring can be accomplished with extreme human feedback efficiency and that labels generated with a mere few minutes of human feedback are sufficient. Code available at: https://github.com/tetrzim/diffusion-human-feedback.