Our method supports training-free acceleration for both latent- and pixel-space image generation models, improving efficiency while preserving quality.
We also introduce a principled fine-tuning framework to push efficiency and quality even further.
[4/7]