RoPEMover: Depth-Aware Object Relocation via Positional Embeddings

How can we perform precise, geometry-consistent object relocation in a single image while maintaining scene-level coherence, such as correct occlusion handling, shadow updates, and plausible background completion, without relying on massive video-based training data?

The researchers propose RoPEMover, a method that treats the internal positional representations of a pretrained diffusion transformer as a manipulable spatial field. By warping the rotary positional embeddings (RoPE) of tokens within an object's mask, the model can induce controlled spatial displacement. To ensure geometric consistency, the authors extend this to a 3D formulation by injecting depth information into the unused temporal axis of the transformer's RoPE. This allows the model to understand the 3D structure of the scene, enabling it to correctly place objects in front of or behind other elements. The model is trained using a combination of synthetic data (from the CLEVR dataset) and a small set of real-world image pairs, utilizing parameter-efficient fine-tuning (LoRA) to adapt the pretrained model to motion tasks.

RoPEMover achieves state-of-the-art performance on standard object motion benchmarks, outperforming existing methods in both visual fidelity and geometric consistency. The model successfully preserves object identity during large displacements and generates plausible content for newly revealed regions. By explicitly modeling depth, the system ensures that shadows and reflections are updated to match the new object position, effectively solving the "copy-paste" artifacts common in earlier editing techniques.

This work demonstrates that powerful spatial and geometric reasoning can be unlocked in existing diffusion transformers by manipulating their internal positional encodings rather than training massive new models from scratch. This provides a data-efficient path toward high-quality, interactive image editing tools that respect the 3D structure of the underlying scene, making it a significant step forward for content creation and visual effects.