SAEdit: Token-Level Control for Continuous Image Editing via Sparse Autoencoder

Abstract

Large-scale text-to-image di_x001B_usion models have become thebackbone of modern image editing, yet text prompts alone do not o_x001B_eradequate control over the editing process. Two properties are especiallydesirable: disentanglement, where changing one attribute does not un-intentionally alter others, and continuous control, where the strengthof an edit can be smoothly adjusted. We introduce a method for dis-entangled and continuous editing through token-level manipulation oftext embeddings. The edits are applied by manipulating the embeddingsalong carefully chosen directions, which control the strength of the tar-get attribute. To identify such directions, we employ a Sparse Autoen-coder (SAE), whose sparse latent space exposes semantically isolateddimensions. Our method operates directly on text embeddings withoutmodifying the di_x001B_usion process, making it model agnostic and broadlyapplicable to various image synthesis backbones. Experiments show thatit enables intuitive and e_x001E_cient manipulations with continuous controlacross diverse attributes and domains.