DynaDrag: Dynamic Drag-Style Image Editing by Motion Prediction

TL;DR

DynaDrag tackles the instability and artifact-ridden edits of prior drag-style methods by introducing a predict-and-move framework that decouples movement prediction from supervision. Motion Prediction proactively forecasts intermediate handle-point movements, while Motion Supervision refines latent diffusion states, with dynamic valid-point selection improving robustness and authenticity. The approach leverages LoRA-finetuned diffusion, DDIM inversion, and a KV-replacement denoising strategy to preserve original content while enabling precise edits. Across face and human datasets, DynaDrag demonstrates higher editability and image fidelity than strong baselines, highlighting its practical potential for pixel-level editing in diffusion-based pipelines.

Abstract

To achieve pixel-level image manipulation, drag-style image editing which edits images using points or trajectories as conditions is attracting widespread attention. Most previous methods follow move-and-track framework, in which miss tracking and ambiguous tracking are unavoidable challenging issues. Other methods under different frameworks suffer from various problems like the huge gap between source image and target edited image as well as unreasonable intermediate point which can lead to low editability. To avoid these problems, we propose DynaDrag, the first dragging method under predict-and-move framework. In DynaDrag, Motion Prediction and Motion Supervision are performed iteratively. In each iteration, Motion Prediction first predicts where the handle points should move, and then Motion Supervision drags them accordingly. We also propose to dynamically adjust the valid handle points to further improve the performance. Experiments on face and human datasets showcase the superiority over previous works.

Figures (11)

• Figure 1: Illustration of existing frameworks, their drawbacks and our framework. Details in these frameworks are omitted. MS in (b,c) represents Motion Supervision. $h_i^k$ is the $i$-th handle point at the $k$-th iteration and $p_i$ represents for its corresponding target point. $p_i^{k+1}$ is the optimization target point at $k$-th Motion Supervision iteration, $\hat{h}_i^{k+1}$ is point tracking searched new handle point whose real location should be $h_i^{k+1}$.
• Figure 2: Details and whole pipeline of our method. $h_i^{k}$ in (c) means the $i$-th handle point at the $k$-th iteration, $p_i$ represents for the $i$-th target point while $f_{h_i^{k}}$ is the feature vector at $h_i^{k}$.
• Figure 3: Dynamic Selection strategy. The feature vector of handle point and the feature vector of its corresponding target point are extracted. Cosine similarity between these two feature vector is calculated. Only the pairs with similarity lower than 0.6 are retained.
• Figure 4: Generalization Study on Motion Prediction. Motion Prediction trained on FaceForensics++ dataset rossler2019faceforensics++ is used to test on DragBench.
• Figure 5: In each iteration, the blue points represent the handle points, the red points denote the target points, the green points indicate the intermediate points in some of the previous iteration steps (to better illustrate the trajectory, we did not plot all the intermediate points), and the yellow line represents the trajectory formed by the intermediate points. It can be observed that the trajectory of the intermediate points does not form a straight line, as would be the case in a typical move-and-track framework. Instead, Motion Prediction allows the handle points to move towards the target points in a smoother and more natural manner (the trajectory generated by Motion Prediction more closely aligns with the actual video trajectory, thereby reducing the difficulty of latent code editing. As a result, the transition between latent codes becomes more natural and smooth).