Training-Free Reward-Guided Image Editing via Trajectory Optimal Control
A new framework for training-free, reward-guided image editing has been introduced by researchers. The method formulates editing as a trajectory optimal control problem, treating the reverse diffusion process as a controllable trajectory from the source image. Adjoint states are iteratively updated to steer editing toward a target reward while preserving semantic content. The approach requires no additional training and is validated across distinct editing tasks. The paper is available on arXiv under ID 2509.25845.
Key facts
- Framework is training-free and reward-guided.
- Editing is formulated as a trajectory optimal control problem.
- Reverse process of diffusion model is treated as controllable trajectory.
- Adjoint states are iteratively updated to steer editing.
- Preserves semantic content of source image while enhancing target reward.
- Validated across distinct editing tasks.
- Paper available on arXiv: 2509.25845.
- Published in 2025.
Entities
Institutions
- arXiv