DragBench: Interactive Image Editing Benchmark

• DragBench is a benchmark dataset and protocol featuring paired drag instructions and editable region masks for precise, user-guided image editing.
• It employs quantitative metrics like Image Fidelity (IF) and Mean Distance (MD) to assess spatial controllability and perceptual realism.
• DragBench has become the de facto standard for evaluating drag-based editing methods across GAN and diffusion models in varied visual contexts.

DragBench is a benchmark dataset and evaluation protocol specifically designed for interactive point-based (“drag-based”) image editing. It provides a systematic and diversified testbed for assessing the spatial controllability, identity preservation, and overall realism of editing techniques where users manipulate specific image regions by dragging handle points to target locations. DragBench has become the de facto standard for quantitative and qualitative assessment in recent literature on drag-based editing with GANs and diffusion models.

1. Motivation and Dataset Structure

The lack of rigorous, diverse, and standardized benchmarks for interactive, localizable image editing motivated the introduction of DragBench (Shi et al., 2023). Previous evaluation settings were either adapted from general image manipulation tasks or constrained to GAN-generated images, with limited focus on spatially explicit, user-guided transformations. DragBench is structured to address these limitations by providing:

• A broad image collection: Multiple categories spanning animals, art, cityscapes, interiors, portraits, landscapes, and objects, including both real photographs and synthetic content. The diversity encompasses scenes with multiple objects, varied styles, and fine-grained local details.
• Paired drag instructions: Each image includes one or more pairs of “handle” and “target” points, defining explicit semantic transformations such as moving a limb, facial feature, or object part to a new spatial location.
• Editable region masks: Binary masks associated with each drag pair, specifying which region of the image is permitted to be manipulated. This ensures edits are localized and spatially constrained.

By covering a multitude of scenarios—from subtle local deformations to complex multi-object transformations—DragBench is capable of exposing weaknesses and strengths in both spatial control and fidelity preservation under challenging, real-world conditions.

2. Evaluation Metrics and Measurement Protocol

DragBench introduced a two-pronged, quantitative evaluation:

Metric	Definition	Desired Value
Image Fidelity (IF)	$$\mathrm{IF} = 1 - \operatorname{LPIPS}(I_\mathrm{orig}, I_\mathrm{edit})$$; assesses perceptual similarity, higher is better.	High
Mean Distance (MD)	Average Euclidean distance between edited “handle” points and specified “target” points across drag pairs. Lower is better.	Low

• Image Fidelity (IF) uses the LPIPS perceptual metric, quantifying how well the edit preserves the overall appearance, identity, and fine details of the original image.
• Mean Distance (MD) evaluates spatial controllability by measuring how closely the semantic content at each handle point is relocated to the user-specified target, as determined by point tracking techniques such as DIFT.

These metrics are jointly reported to ensure that models are not only precise in relocating points but also restrained in altering the global appearance, thereby discouraging trivial solutions that ignore either realism or control.

3. Positioning within the Landscape of Editing Benchmarks

DragBench represents a substantial departure from prior GAN-based editing tests, which either lacked spatially localized instructions or only considered the latent spaces of generative models. Unlike previous benchmarks, DragBench is designed from the ground up for explicit, point-wise transformations and supports both GAN and diffusion-based pipelines (Shi et al., 2023).

The dataset’s comprehensive coverage of categories, paired dragging instructions, and region masks, combined with its strict dual-metric evaluation, creates uniquely demanding scenarios. These characteristics have directly influenced the development of advanced editing strategies explicitly targeting the trade-off between spatial accuracy and preservation of semantic content. As a result, DragBench remains the principal protocol for empirical validation of new drag-based editing methods (Cui et al., 7 Mar 2024, Zhao et al., 24 May 2024, Yin et al., 15 Sep 2025).

4. Integration in Drag-based Editing Model Development

DragBench has become integral to the assessment of both foundational and state-of-the-art editing methods:

• DragDiffusion (Shi et al., 2023): Extends DragGAN's framework to large-scale diffusion models and uses DragBench for direct comparison. Ablation studies leveraging DragBench revealed the impact of loss functions, feature map selection, and inversion strategy on the spatial control/fidelity tradeoff.
• StableDrag (Cui et al., 7 Mar 2024): Tackles inaccurate point tracking and motion supervision shortcomings exposed by DragBench, introducing discriminative point tracking and confidence-based supervision; quantitative enhancements on DragBench (e.g., IF and MD) validate the principled improvements.
• FastDrag (Zhao et al., 24 May 2024): Employs a latent warpage function for one-step editing, demonstrating (on DragBench) substantial reductions in computation time and strong trade-offs in IF/MD compared to prior multi-iteration approaches.
• LazyDrag (Yin et al., 15 Sep 2025): Leveraging MM-DiTs, explicit correspondence maps, and region-based attention control for robust, TTO-free, and high-fidelity edits, with DragBench evaluations showing lower MD and higher VIEScore (semantic and perceptual metrics) than existing methods.

DragBench also supports in-depth ablation and variant evaluations, such as testing the stability of tracking, the effect of different inversion step counts, or the role of additional denoising and attention mechanisms for identity preservation.

5. Empirical Results and Benchmarking Outcomes

DragBench’s quantitative benchmarks have consistently established a hierarchy among competing methods, pushing the field toward methods excelling along both spatial and semantic axes. Typical empirical findings include:

• Consistent improvement: Newer, geometry-aware, or more sophisticated point-tracking approaches (e.g., StableDrag, FastDrag, FlowDrag) show lower MD (higher spatial accuracy) and higher IF (greater fidelity) compared to their GAN or vanilla diffusion-based predecessors.
• Trade-off elucidation: Side-by-side scatter plots (MD vs IF) reveal that approaches achieving low MD at the expense of IF are penalized, as are models that yield sharp images but fail to align content with specified targets.
• Efficiency studies: FastDrag demonstrates that, relative to DragDiffusion (~60 sec/point), its architecture reduces the average processing time to ~3 sec/point on DragBench, at comparable or superior MD and IF.
• User and model-based evaluation: LazyDrag supplements conventional metrics with VIEScore (semantic consistency, perceptual quality, and overall scores from multi-modal evaluators like GPT-4o), confirming advantages in both objective and subjective quality dimensions.

6. Limitations and Directions for Extensions

Although DragBench provides a critical foundation for evaluating drag-based editing, certain limitations are acknowledged (Shi et al., 2023, Koo et al., 11 Jul 2025, Yin et al., 15 Sep 2025):

• Lack of ground truth: DragBench does not provide canonical, ground-truth images for the post-edit state; evaluation is restricted to point placement and perceptual similarity, rather than direct pixel-wise comparison. This is remedied in subsequent datasets such as VFD-Bench (Koo et al., 11 Jul 2025), where temporally adjacent video frames provide explicit before/after pairs.
• Disentanglement of ambiguity: When drag instructions are under-defined or ambiguous (e.g., moving a limb with multiple possible plausible endpoints), DragBench does not resolve a unique semantic interpretation.
• Resolution of fine-grained assessment: Side effects such as subtle artifacts and context-aware semantic transformations may not be fully captured by point-based MD, IF, or LPIPS alone, spurring the introduction of multi-modal and prompt-based evaluation metrics (e.g., VIEScore).

A plausible implication is that the field is moving toward benchmarks that incorporate richer ground-truth targets, fine-grained region annotation, and integrated semantic evaluators to further stress-test spatial and semantic controllability in increasingly unconstrained environments.

7. Impact and Future Prospects

DragBench has shaped the research landscape not only as a technical benchmark but also as a factor driving methodological innovation. Its rigorous constraints have spurred new directions, including 3D geometry-aware editing (FlowDrag (Koo et al., 11 Jul 2025)), multi-modal and text-guided manipulations (LazyDrag (Yin et al., 15 Sep 2025)), as well as cascaded and multi-round editing workflows.

Significantly, DragBench's design principles and evaluation metrics continue to influence the development of both datasets (such as Drag100 and VFD-Bench) and metrics (e.g., DAI, VIEScore, Gemini Score) that aim to marry spatial precision, perceptual realism, and semantic compliance. Continued expansion along these axes is expected to further enhance the reliability, generality, and practical relevance of point-based image editing technologies.