LatentPS: Image Editing Using Latent Representations in Diffusion Models
Zilong Wu, Hideki Murata, Nayu Takahashi, Qiyu Wu, Yoshimasa Tsuruoka · 2025
Generative image models have significantly advanced recently, enabling detailed image generation from textual prompts. However, accurately describing complex image with textual prompts alone is often challenging. To address this, we propose LatentPS, a novel method for image editing that modifies latent representations during the intermediate stages of the diffusion process. LatentPS leverages the properties of the diffusion process to maintain the structure of synthesized images while editing images through simple operations on latent representations directly. We show that LatentPS achieves promising results in various image editing tasks including moving, resizing and pasting.