GoodDrag presents a new methodology aimed at refining drag editing processes, addressing prevalent issues such as distortion and accumulated perturbations. The framework emphasizes two pivotal practices: Alternate Drag and Denoising, and Information-Preserving Motion Supervision. Leveraging diffusion models renowned for their capacity in generating high-fidelity images, GoodDrag orchestrates forward and reverse processes to enhance image stability and fidelity.

Incorporating the Alternate Drag and Denoising practice, GoodDrag strategically employs motion supervision and denoising techniques to iteratively refine image samples. By mitigating distortions and preserving original features, this approach endeavors to elevate the quality of edited images. Furthermore, the framework introduces Information-Preserving Motion Supervision, ensuring precise content transfer during drag editing tasks by meticulously updating samples and tracking points.

To facilitate rigorous evaluation, GoodDrag introduces the Drag100 dataset, comprising diverse images annotated with masks and control points for controlled assessments of drag editing algorithms. Encompassing various drag tasks including relocation, rotation, rescaling, content removal, and creation, Drag100 provides a comprehensive benchmark for assessing algorithmic performance. Additionally, two novel quality assessment metrics, Dragging Accuracy Index (DAI) and Gemini Score (GScore), are proposed to quantitatively gauge the efficacy of drag editing approaches in semantic content transfer and achieving high-quality edits.