PhyEditBench: A Real-World Multi-Stage Benchmark for Physics-Aware Image Editing

2026-06-24

AI Summary

This work addresses the lack of systematic evaluation of physical consistency in existing image editing benchmarks by introducing the first physics-aware assessment framework. The authors construct a multi-stage benchmark dataset derived from real-world videos, comprising 238 high-resolution authentic samples and 35 physically implausible synthetic instances, organized under a four-level hierarchical taxonomy to comprehensively evaluate models’ physical reasoning capabilities. They further propose PhyWorld, a training-free baseline method that leverages the video generation process as an inference mechanism and incorporates a latent-space compression strategy during inference to enable physics-aware editing. Experimental results reveal significant deficiencies in the physical consistency of current state-of-the-art models, while PhyWorld—without any training—outperforms existing approaches, demonstrating the effectiveness of leveraging dynamic visual information from videos to enhance the plausibility of image edits.

Abstract

While instruction-based image editing, enabled by multi-modal generative models, has advanced significantly, existing benchmarks lack a comprehensive evaluation of physics-based reasoning, a critical capability for handling real-world scenarios. To address this, we introduce PhyEditBench, a benchmark designed to assess the physical understanding of editing models. Guided by a hierarchical taxonomy, we establish 4 primary classes and 12 subclasses. It comprises 238 high-quality, high-resolution, real-world instances meticulously extracted from videos to capture authentic physical dynamics, alongside 35 synthetic Anti-Physics instances. Our empirical analysis of current SOTA editing methods exposes substantial limitations in their physics-based reasoning. We further propose a training-free baseline named PhyWorld that uses test-time scaling and a latent reduction strategy. PhyWorld outperforms comparable models and suggests that the video generation process can effectively serve as a reasoning mechanism for image editing. The project page is available at https://github.com/Previsior/PhyEditBench.

Problem

• physics-aware image editing
• real-world benchmark
• physical reasoning
• image editing evaluation
• multi-stage editing

Innovation

• physics-aware image editing
• real-world benchmark
• multi-stage evaluation
• test-time scaling
• video-based reasoning