Reproduce “Efficient Text-to-Image Generation: An Adaptive Step Schedule Controller for Diffusion Models” 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
Kuluhan Binici, Cihan Acar, Shivam Aggarwal, Siying Liu, Tulika Mitra
ADSC is a training-free controller that adapts diffusion sampling effort to each prompt by switching from fine to coarser denoising schedules when classifier-free guidance signals indicate convergence. In Stable Diffusion v1-5 experiments, it maintained text alignment while improving reference-image similarity over fixed schedules at comparable average step counts, with the clearest gains obtained using DDIM.
Key insights
- Prompt-conditioned guidance dynamics can serve as a runtime signal for deciding when additional fine-grained denoising is no longer necessary.
- Adaptive schedules allocate more steps to complex prompts and fewer steps to simpler scenes, rather than imposing a uniform computational budget.
- Performance depends on the sampler and dataset: the strongest and most consistent improvements appear with DDIM, while PNDM results are more mixed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.