Region-based Dragging Benchmark (ReD Bench)

• The paper introduces ReD Bench as a standardized testbed for evaluating interactive image editing, leveraging region masks, centroids, and explicit operation classes.
• It details a hand-curated dataset of 120 diverse photographs paired with structured JSON annotations for relocation, deformation, and rotation tasks.
• Evaluation protocols use dual metrics—Image Fidelity and Mean Distance—to comprehensively assess both perceptual quality and geometric precision in drag editing.

Region-based Dragging Benchmark (ReD Bench) rigorously defines a standardized testbed for evaluating interactive image editing systems that perform fine-grained object movement, deformation, or rotation via region-aware dragging instructions. Unlike earlier point-based benchmarks, ReD Bench centers on annotated region masks, spatial centroids, and explicit operation classes, enabling systematic, discriminative assessment of both geometric and perceptual editing fidelity under diverse manipulation scenarios (Zhou et al., 2 Oct 2025, Zafarani et al., 13 Dec 2025).

1. Dataset Construction and Structure

ReD Bench is composed of 120 hand-curated real photographs spanning diverse content domains, with an emphasis on manipulation complexity:

• Scenes: Indoors, outdoors, aerial/coastal, people, and objects.
• Tasks: Each image is paired with one or more region-level dragging operations from three main classes: relocation, deformation, and rotation.
• Selection: Images and operations were manually sampled to balance small/large, rigid/nonrigid regions, and single/multiple simultaneous drags.
• Intended usage: The benchmark is released as a held-out evaluation set with no train/val/test splits, supporting zero-shot assessment of interactive drag-and-drop editors (Zhou et al., 2 Oct 2025).

For each operation, annotators traced a source region mask (PNG), calculated its centroid $b_k = (1/|M_k|)\sum_{p\in M_k} p$, and provided a target point $t_k$. Rotation tasks include an anchor point $a_k$. All annotation masks and labels underwent author-side consistency checks.

2. Annotation Format and Instruction Schema

Each ReD Bench sample is supplied as a structured JSON instance containing:

• region_operations: For each instance, specifies task (relocation, deformation, rotation), two centroids ([source, target]), and (optionally) an anchor point for rotations.
• point_operations: Packaged for direct comparison with handle-point-based models, gives [begin_points] and [target_points] for traditional point-dragging.
• background_prompt: A natural-language summary of the scene context.
• editing_prompt: A free-form textual instruction describing the intended edit.

All region masks and centroids are included to facilitate reproducibility, eliminating ambiguity from algorithmic segmentation. No formal inter-annotator agreement is reported, but manual verification ensures mask-label alignment.

3. Evaluation Protocol and Region-based Metrics

ReD Bench implements a dual-metric regime addressing both geometric and perceptual performance.

Image Fidelity (IF), based on the LPIPS perceptual similarity (lower is better, but converted here as $1-\mathrm{LPIPS}$ so higher is better):

• $IF_{s2t}$: Fidelity of source content after affine-warping to the target region.
• $IF_{s2s}$: Measures the degree to which the source region is "cleared out" from the original location.
• $IF_{bg}$: Assesses background consistency outside edited regions using an adaptive background mask $B$.

Mean Distance (MD):

• $MD_1$: Masked region mean distance, as in DragLoRA, quantifies L2 difference between feature centroids pre- and post-edit (lower is better).
• $MD_2$: Centroid-based feature region distance, per RegionDrag, eliminates manual correspondences in favor of automatic centroid feature matching.

Reporting of both MD and IF metrics, calculated with provided region masks and centroids, enables robust, reproducible comparison of editing algorithms (Zhou et al., 2 Oct 2025).

4. Baselines and Experimental Setup

Zero-shot inference runs on ReD Bench have been conducted for a range of region-based and point-based image drag editors:

• Region-based: RegionDrag, FastDrag (region only, no optimization), InstantDrag, DragLoRA, FreeDrag, DragNoise, GoodDrag, CLIPDrag, DragDiffusion.
• State-of-the-art: DragFlow (FLUX.1-dev with FireFlow inversion and adapter-enhanced inversion), notable for region-level affine supervision, strict background constraints, and InstantCharacter adapter for subject consistency.

Ablation experiments evaluate incremental influence of: (a) point-based FLUX, (b) region-level affine, (c) background preservation, (d) adapter inversion.

Quantitative Results (ReD Bench, Table 1) (Zhou et al., 2 Oct 2025):

Method	MD₁ (↓)	MD₂ (↓)	IF₍s2t₎ (↑)	IF₍s2s₎ (↑)	IF₍bg₎ (↑)
DragFlow	19.46	4.48	0.958	0.934	0.992
GoodDrag	20.38	4.50	lower	lower	lower

DragFlow demonstrates the best geometric and perceptual scores, reflecting the advantage of region-dependent affine constraints and adapter-based personalization. Qualitative outcomes confirm superior structural preservation and minimal background distortion over prior point-oriented methods.

5. Benchmark Design Rationale and Comparisons

ReD Bench’s emphasis on region masks and centroids distinguishes it from preceding datasets that rely on sparse handle-point manipulation or lack real paired ground truth images (as in RealDrag (Zafarani et al., 13 Dec 2025)). By anchoring fidelity and alignment computation to the explicitly supplied region, ReD Bench resolves ambiguities in spatial correspondence and supports direct evaluation of complex edits (multi-point, articulated, or non-rigid transformations).

Standardized instruction format—combining JSON operations, mask data, and natural-language prompts—facilitates head-to-head comparison across model families (GAN, diffusion, optimization-based, and hybrid). The inclusion of both region and point correspondences enables benchmarking of legacy and advanced semantic drag methods under uniform evaluation criteria.

6. Metric Interpretation, Best Practices, and Impact

ReD Bench’s dual-metric system explicitly disentangles spatial alignment (MD) from perceptual and content fidelity (IF family), overcoming limitations of pure pixel-level measures. This protocol incentivizes comprehensive model development that balances geometric precision, structural region integrity, and background preservation.

Empirical results reveal that methods optimizing only for point-accuracy often introduce non-local distortions or background artifacts, while region-based supervision—especially with adapters and hard background constraints—significantly improves both task and perceptual outcomes.

For future use, reporting all region-aware metrics jointly is recommended, employing editable region masks for both test/validation. Task-wise breakdown (relocation, deformation, rotation) should accompany aggregate results to expose algorithmic strengths, weaknesses, and trade-offs on different manipulation primitives.

7. Significance for Interactive Image Editing Research

ReD Bench formalizes the evaluation of region-based drag editing under rigorous, reproducible, and discriminative conditions. Its design and metrics have already enabled the identification of failure modes in both classical and recent drag-editing models, guiding diagnostic analysis and ablation studies (Zhou et al., 2 Oct 2025). By bridging region-matched geometry with perceptual quality, ReD Bench underpins progress toward controllable, high-fidelity, and semantically aligned image editing systems that generalize to complex and multi-object scenarios.