RoPEMover: Depth-Aware Object Relocation via Positional Embeddings

A new image-editing method called RoPEMover can relocate objects within a single photograph while preserving scene-level consistency such as shadows, reflections, and occlusions, according to a paper submitted to arXiv on 25 June 2026 [1]. The approach operates directly on the positional representations of diffusion transformers, the class of generative models that underpin many modern image-synthesis systems [1]. The researchers exploit rotary positional embeddings, or RoPE, which define a structured spatial field that can be explicitly manipulated to induce controlled motion [1]. By warping the RoPE encodings of masked tokens according to a user-specified drag signal — encoding the desired displacement and scale change — the method directly induces the intended spatial transformation within the attention layers of the diffusion transformer [3]. A key technical contribution is the extension of standard 2D RoPE into a depth-aware formulation that encodes 3D spatial structure [1]. The system estimates a metric depth map from the input image, modifies it based on the desired object displacement, and injects it into the otherwise unused temporal axis of a factorized 3D RoPE [3]. This provides the model with explicit geometric context about the scene, enabling correct occlusion handling, plausible completion of newly revealed regions, and consistent updates to scene-dependent effects such as shadows and illumination [4]. The model is trained using synthetic data combined with a small set of real images via parameter-efficient fine-tuning [1]. Despite minimal real supervision, it preserves object identity under large spatial displacements [1]. Experimental results on standard object motion benchmarks demonstrate state-of-the-art performance across all evaluation metrics [2]. The use of positional embeddings to encode geometric information is an active area of research. A separate 2024 study proposed object-wise position embedding to inject depth information into transformer decoders for multi-view 3D object detection, improving 3D-aware feature generation [5]. More recently, a 2025 technical report introduced Multimodal Rotary Position Embedding, decomposing rotary embeddings into temporal, height, and width components to improve camera-only 3D object detection and velocity estimation in autonomous driving scenarios [6]. RoPEMover extends this line of inquiry into the domain of single-image editing, repurposing the temporal axis as a depth channel rather than a time dimension [3]. The paper was posted on arXiv, the open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and hosts more than two million articles across physics, computer science, and related fields [10].