AdvDMD Unifies DMD Distillation and RL for Few-Step Generation
AdvDMD is an innovative technique that merges Distribution Matching Distillation (DMD) with reinforcement learning to enhance the quality of image generation over few steps. Unlike traditional methods that complicate the process by combining RL and distillation, AdvDMD utilizes the adversarially trained discriminator from DMD2 as its reward model, giving low scores to generated images and high scores to authentic ones. This approach is trained on both intermediate and final outputs, with the goal of outperforming the teacher model while requiring fewer sampling steps.
Key facts
- AdvDMD unifies DMD distillation and reinforcement learning
- Uses adversarially trained discriminator from DMD2 as reward model
- Assigns low scores to generated images, high scores to real ones
- Trained on both intermediate and final outputs
- Aims to surpass teacher model performance with fewer steps
- Addresses performance degradation in few-step generation
- Simplifies existing combinatorial approaches
- Published on arXiv (2604.28126)
Entities
Institutions
- arXiv