Reproduce “Target-Driven Distillation: Consistency Distillation with Target Timestep Selection and Decoupled Guidance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cunzheng Wang, Ziyuan Guo, Yuxuan Duan, Huaxia Li, Nemo Chen, Xu Tang, Yao Hu
Target-Driven Distillation (TDD) enhances the efficiency and output quality of consistency distillation for diffusion models through a refined target timestep selection strategy, decoupled guidance, and optional sampling improvements. The method yields competitive FID scores and high human preference ratings for few-step image generation.
Key insights
- Optimizing target timestep selection to focus on predictions relevant to expected few-step inference schedules can significantly improve training efficiency and the quality of generated images.
- Decoupling the guidance scale during distillation from inference-time application provides greater user control, allowing for dynamic adjustments to image characteristics without model retraining.
- Strategic use of sampling enhancements like x0 clipping, especially in early denoising steps, can effectively mitigate defects and improve realism in few-step generated images without introducing excessive contrast.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.