SREdit-Bench: A Benchmark for Spatial Reasoning in Image Editing

Dataset Description

SREdit-Bench is a challenging benchmark designed to comprehensively evaluate the spatial reasoning capabilities of image editing models. While existing benchmarks primarily focus on object-salient scenes, they often under-represent spatially complex editing scenarios (e.g., multiple similar editable entities, non-object-salient scenes, and fine-grained object referral).

To fill this gap, we introduce SREdit-Bench, comprising 590 carefully curated high-resolution (>1024×1024) samples. It focuses on evaluating a model's true spatial reasoning ability under complex editing scenarios

Key Features

1. Sophisticated Scenes: The benchmark deliberately includes images with multiple entities and non-object-centric layouts to test robust localization.
2. Fine-grained Target Referral: The instructions require deep spatial and contextual understanding, categorized into three modes:
   • Spatial: Instructions with explicit location or relational cues.
   • Property: Instructions using appearance or attribute-based descriptions.
   • Knowledge: Implicit, context-dependent cues that require background or commonsense knowledge.

Dataset Structure

(Note: Please adjust the field names based on your actual Hugging Face dataset columns)

Each instance in the dataset typically contains:

• image: The source high-resolution image.
• instruction: The editing instruction containing fine-grained referring expressions.
• referral_type: The category of the instruction (Spatial, Property, or Knowledge).
• edit_type: The category of the editing task (e.g., Add, Remove, Replace, Modify).

Evaluation Protocol

We follow the VIEScore, we evaluate the edited images across three main metrics:

1. Semantic Consistency (SC): Measures how well the edited result faithfully follows the given instruction.
2. Perceptual Quality (PQ): Evaluates the image's naturalness and the presence of visual artifacts.
3. Overall (O): The geometric mean of SC and PQ, averaged across all test samples.

Usage

You can easily load this dataset using the datasets library:

from datasets import load_dataset

dataset = load_dataset("zjYinnnn/SREdit-Bench")
print(dataset['test'][0])