GeoNaviDiff: Transferring Disentangled Semantic Controls to Diffusion Features via Local Riemannian Navigation

Key points

• Disentangled real-image manipulation requires factor-specific semantic control while preserving image fidelity and non-target content.1
• StyleSpace $S$ provides interpretable controls but is limited by GAN inversion, whereas diffusion models offer stronger image realization without an explicit factorized control basis.1
• 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.1
• 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.1

Sources (1)

• [1]GeoNaviDiff: Transferring Disentangled Semantic Controls to Diffusion Features via Local Riemannian Navigation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 09:52 AM

  We introduce GeoNaviDiff, an unsupervised cross-space framework that transfers disentangled StyleSpace interventions to local Riemannian trajectories in diffusion $h$-space.

  Disentangled real-image manipulation requires factor-specific semantic control while preserving image fidelity and non-target content.

Extractive summary: sentences quoted from the sources.