Training-free multiview editing with localized control + cross-view consistency.
Authors: Mohamed Gallai, Abby Stylianou
Venue: Applied Imagery Pattern Recognition Workshop (AIPR 2025) — to appear in Springer proceedings
🏆 Best Student Paper — AIPR 2025 (Oral)
📄 Preprint (author-created): paper/preprint.pdf
Multiview image editing (wide-baseline pairs) often fails due to cross-view inconsistency and poor localized control.
This work introduces a lightweight, training-free pipeline that improves consistency while keeping edits localized.
- Mask-guided partial diffusion to localize edits
- Training-free cross-view feature sharing to improve consistency across viewpoints
High-level pipeline of the proposed imgs2imgs framework.
Two-pass inference strategy used to transfer structural features between views.
Comparison against baseline methods showing improved consistency.
Example of localized, consistent multiview editing result.
🏆 Best Student Paper — AIPR 2025 (Oral)
imgs2imgs: Improving Visual Consistency in Multiview Image Editing
Mohamed Gallai, Abby Stylianou
Applied Imagery Pattern Recognition Workshop (AIPR), 2025
To appear in Springer proceedings