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

Instruction-based image editing has evolved from simple style transfers to complex semantic manipulations. However, most existing benchmarks focus on spatial layout or attribute binding, failing to test whether models understand the underlying physical laws—such as gravity, fluid dynamics, or material deformation—that govern how objects change over time. This paper introduces PhyEditBench to address this critical gap, providing a rigorous testbed for evaluating how well generative models simulate real-world physical transitions.

PhyEditBench consists of 238 high-resolution, real-world instances extracted from videos, categorized into a hierarchical taxonomy of 12 subclasses across four primary domains: Deformation & Fracture, Fluid Dynamics, Rigid Body & Interaction, and State Change & Environment. To ensure robust evaluation, the benchmark also includes 35 synthetic Anti-Physics instances that force models to ignore common visual habits (like gravity pulling objects down) and instead follow explicit counterfactual instructions. Each instance provides a multi-stage trajectory, allowing researchers to evaluate not just the final result, but the intermediate steps of a physical process.

To overcome the limitations of static editing models, the authors propose PhyWorld, a training-free framework that treats image editing as a temporal generation task. By utilizing pretrained video generation models, PhyWorld interprets the generation of intermediate frames as an implicit reasoning process. The framework incorporates Test-Time Scaling (TTS) and a latent reduction strategy, which allows the model to iteratively optimize the output against a video reward model. This approach ensures that the final edited image is not only semantically aligned with the user's prompt but also physically plausible.