GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models

Xiangyu Xu

TL;DR

This paper introduces GoodDrag, a novel framework for drag-based image editing using diffusion models. It tackles instability and distortion issues in existing methods by alternating drag and denoising steps (AIDD) and incorporating an information-preserving motion supervision to maintain feature fidelity. The authors also introduce a new benchmark dataset (Drag100) and evaluation metrics (DAI, GScore) to advance drag editing research.

What Problem Does This Paper Address?

Drag editing allows users to intuitively manipulate images by specifying start and end points, as if "dragging" an object. While early methods relied on GANs, they struggled with generic images. More recent approaches leverage diffusion models, which offer better image quality and flexibility. However, these diffusion-based methods often suffer from instability, leading to distorted outputs and a failure to precisely adhere to control points. This instability is partly due to the accumulation of perturbations during the editing process, especially when all drag operations are performed at once before denoising. Additionally, existing methods can experience "feature drifting," where the visual characteristics of the dragged object deviate from its original appearance, causing artifacts.

Methods

GoodDrag introduces two key innovations:

  • Alternating Drag and Denoising (AIDD): Instead of performing all drag operations and then denoising, AIDD interleaves drag and denoising steps. This approach breaks down the editing process into smaller, more manageable chunks, preventing the accumulation of large perturbations and leading to more stable and higher-fidelity results. The authors illustrate this with a visual trajectory showing a more gradual movement towards the desired outcome compared to the abrupt changes in existing methods.
  • Information-Preserving Motion Supervision: This technique addresses the "feature drifting" problem. It ensures that the features of the dragged object remain consistent with its original appearance throughout the editing process. This is achieved by modifying the motion supervision loss to explicitly preserve the original features of the starting point, leading to more accurate manipulation and reduced artifacts.

The overall pipeline involves DDIM inversion to get a latent representation, followed by the AIDD process (alternating drag and denoising steps), and finally, point tracking to ensure precise handle point localization.

Datasets

  • Drag100 Dataset: The authors introduce a new benchmark dataset called Drag100, specifically designed for drag editing evaluation. It contains 100 images (85 real, 15 AI-generated) with carefully labeled masks and control points. This dataset aims to provide a controlled environment for evaluating different drag editing algorithms, unlike previous datasets that lacked explicit masks. Drag100 also covers a diverse range of tasks, including relocation, rotation, rescaling, content removal, and content creation, and features various categories like animals, landscapes, and objects.

Key Insights and Ideas

  • Incremental Editing is Key: The core insight is that addressing perturbations incrementally, rather than allowing them to accumulate, significantly improves image editing quality and stability. This is analogous to fixing small problems as they arise rather than waiting for them to become major issues.
  • Feature Drifting is a Major Artifact Source: The paper identifies feature drifting as a primary cause of artifacts in drag editing. Simply encouraging the next handle point to be similar to the current one isn't enough; it needs to be anchored to the original appearance.
  • Diffusion Models Offer a Better Foundation: The paper highlights the inherent advantages of diffusion models for drag editing due to their superior generative capabilities and flexibility compared to GANs.
  • Multimodal Models for Evaluation: Leveraging Large Multimodal Models (LMMs) like Gemini for image quality assessment (GScore) provides a more reliable and human-aligned evaluation than traditional No-Reference Image Quality Assessment (NR-IQA) metrics.

Key Results

  • Quantitative Performance: GoodDrag consistently outperforms state-of-the-art diffusion-based drag editing methods (DragDiffusion, SDE-Drag) in terms of both drag accuracy (DAI) and image quality (GScore) on the Drag100 dataset. For instance, GoodDrag achieves a DAI of 0.0623 with a patch radius of 20, significantly lower than DragDiffusion's 0.1146. Its GScore of 7.94 also surpasses DragDiffusion's 6.90.
  • Qualitative Improvements: Visual comparisons show GoodDrag producing significantly fewer artifacts and more faithful edits compared to existing methods, even when handling complex scenarios and diverse editing tasks. The AIDD framework demonstrably reduces distortions, and information-preserving motion supervision effectively prevents feature drifting.
  • User Study: A user study confirms GoodDrag's superiority, with participants ranking its results higher for both drag accuracy and perceptual quality.

Limitations and Future Work

  • Computational Cost: While GoodDrag improves efficiency by distributing operations, it still requires significant computation. The paper mentions a runtime of about one minute for editing on an A100 GPU for a 512x512 image after LoRA finetuning.
  • Optimization Challenges: The information-preserving motion supervision, while effective, can be more challenging to optimize, sometimes requiring multiple steps within a single drag operation.
  • Future Directions: The authors suggest exploring the integration of GoodDrag with other image editing tasks and extending its capabilities to video editing scenarios. They also highlight the potential for further research into more sophisticated drag editing algorithms.

Summary updated

Abstract

In this paper, we introduce GoodDrag, a novel approach to improve the stability and image quality of drag editing. Unlike existing methods that struggle with accumulated perturbations and often result in distortions, GoodDrag introduces an AlDD framework that alternates between drag and denoising operations within the diffusion process, effectively improving the fidelity of the result. We also propose an information-preserving motion supervision operation that maintains the original features of the starting point for precise manipulation and artifact reduction. In addition, we contribute to the benchmarking of drag editing by introducing a new dataset, Drag100, and developing dedicated quality assessment metrics, Dragging Accuracy Index and Gemini Score, utilizing Large Multimodal Models. Extensive experiments demonstrate that the proposed GoodDrag compares favorably against the state-of-the-art approaches both qualitatively and quantitatively.
BibTeX
@inproceedings{xu2025gooddrag,
  author = {Xiangyu Xu},
  title = {{GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=VtYfbvwpWp},
}

Loading related content...

Figure