ONLS: Optimal Noise Level Search in Diffusion Autoencoders Without Fine-Tuning

Zihan Wang

Keywords: Computer Vision

Abstract: An ideal counterfactual estimation should achieve balance of precise intervention and identity preservation. Recently, Classifier-Guided Diffusion Model is proven effective to produce realistic and minimal counterfactuals. However, perfect intervention is often challenging to find and requires tedious fine-tuning. In this work, we propose Optimal Noise Level Search (ONLS), which leverages statistics from the guidance to automatically capture the balance without any fine-tuning process or extra network design. We demonstrate that our ONLS could accurately identify the optimal noise level for counterfactual estimation. The optimal per-sample results further contribute to an overall performance enhancement across the dataset. Preprocessing, curated dataset, and code are released on our project page: \url{https://github.com/ImNotPrepared/ONLS}.

Submission Number: 17