Reference Conditioned Boundary Transport for Shape Stable One Shot Diffusion Editing of Object Centric Videos with Temporal Priors

• Sagar Adhikari

  Department of Computer Science and Information Technology, Jana Bhawana Campus, Godawari-11, Chapagaun, Lalitpur, Bagmati Province, Nepal

  Author

• Chandresh Kumar Chhatlani

  Department of Computer Science and IT, JRN Rajasthan Vidyapeeth, Airport Road, Pratap Nagar, Udaipur, Rajasthan 313001, India

  Author

Diffusion models have made object level video editing more controllable, yet edits that depend on a reference object still expose a gap between appearance transfer and motion preservation. A video editor must keep framewise geometry coherent while allowing the edited object to depart from the source silhouette. We study this problem through a one shot setting in which a source video, a source prompt, a target prompt, and a single target image are available. The proposed method, Boundary Transport Diffusion, estimates an editable object support by transporting a reference induced boundary field through denoising features, then regularizes the generated latent trajectory with a bidirectional temporal prior. The approach does not train a video diffusion model from scratch and only updates compact temporal adapters on the source clip. We evaluate the method on 96 short videos spanning vehicles, animals, sports objects, footwear, hand held tools, and deformable apparel. Across these clips, the method improves target image alignment by 0.061 DINO similarity over a strong reference conditioned baseline, reduces warped temporal LPIPS by 18\%, and lowers boundary flicker by 27\% relative to mask propagation without boundary transport. Human raters preferred the proposed outputs for reference fidelity in 63\% of paired comparisons and for temporal stability in 58\%. The results suggest that reference conditioned masks should be treated as dynamic latent variables rather than static spatial constraints. Failure cases remain visible for articulated thin structures, severe occlusion, and edits requiring new physical motion not present in the source clip.

• PDF

Copyright (c) 2026 International Journal of Scientific and Applied Computing Research

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.