Geo navidiff improves image edits with semantic controls and diffusion models
GeoNaviDiff: Transferring Disentangled Semantic Controls to Diffusion Features via Local Riemannian Navigation
Computer Vision and Pattern Recognition
Summary
Changing specific parts of a photo without messing up other details is hard. The authors created GeoNaviDiff, a way to mix easier-to-understand controls from one image model with the higher quality of another model. This lets them change images more precisely and smoothly without needing extra training for each change. Their method works better than others in keeping the picture clear and the edits focused.
What this means in practice
- •For graphic designers: Create photo edits that selectively modify features while keeping original image quality intact using a streamlined control method.
- •For mobile app developers: Integrate enhanced image editing features that allow precise adjustments without needing separate models for each edit type.
Authors
Tao Sun, Huanting Guo, Baao Xie, Zhou Fang, Qiuyu Chen, Hao Zheng, Wenjun Zeng
Abstract
Disentangled real-image manipulation requires factor-specific semantic control while preserving image fidelity and non-target content. StyleSpace $S$ provides interpretable controls but is limited by GAN inversion, whereas diffusion models offer stronger image realization without an explicit factorized control basis. We introduce GeoNaviDiff, an unsupervised cross-space framework that transfers disentangled StyleSpace interventions to local Riemannian trajectories in diffusion $h$-space. GeoNaviDiff constructs reconstruction-decoupled supervision by extracting a spherical Log-map tangent from a matched StyleGAN reconstruction-edit pair and parallel-transporting it to the real-image diffusion state. A shared, factor-conditioned Navigator predicts this tangent at inference, while norm-preserving geodesic realization and source-anchored rollout stabilize the reverse process without edited references or factor-specific models. Experiments across multiple datasets show that GeoNaviDiff outperforms the compared methods in fidelity, factor selectivity, edit locality, and trajectory smoothness.