PhysicsEdit: Teaching Image Editing Models to Respect Physics

Command-based photo editing models are impressive for following commands. But when planning involves physical interaction, they often fail to respect the rules of the real world. In their paper “From Statics to Dynamics: Physics-Aware Image Processing for Previous Changes,” the authors present PhysicEdit, a framework that treats image editing as a physical transformation rather than a static transformation between two images. This change improves accuracy in physics-heavy situations.

Failure of AI Image Generation

You create a room with a light and ask the model to turn it off. The lamp goes out, but the light in the room does not change. The shadows are always inconsistent. The order is followed, but the illumination physics is ignored.

Now put the straw in a glass of water. The straw comes out of the glass but stays perfectly straight instead of bending due to resistance. The arrangement looks fine at first, but it violates optical physics. These are the very failures that PhysicsEdit aims to fix.

Also read: Top 7 AI image generators to try in 2026

The Problem with Current Image Editing Models

Most command-based programming models follow a specific setup.

• You provide a source image.
• He gives a planning order.
• The model produces a modified image.

This works well for semantic editing like this:

• Change the color of the shirt to blue
• Substitute a cat for a dog
• Remove the chair

However, this setup treats editing as a static map between two images. It does not simulate the process leading from the initial state to the final state.

This becomes a problem in physics-heavy situations such as:

• Put the straw in a glass of water
• Let the ball fall on the pillow
• Turn off the light
• Freeze a soda can

This planning requires an understanding of how the laws of nature affect the scene over time. Without modeling that change, the system often produces results that look plausible at first glance but break down under closer inspection.

From a static map to a physical change

PhysicsEdit suggests a different layout.

Instead of directly predicting the final image from the source image and command, it treats the command as a visual trigger. The source image represents the original state of the scene. The final image represents the result after the scene has evolved under the laws of nature.

In other words, planning is seen as a problem of regime change rather than direct change.

This distinction is important.

A typical editing dataset only provides the first image and the last image. There are no intermediate steps. As a result, the model learns what the output should look like, but not how the scene should evolve to reach that state.

PhysicsEdit addresses this limitation by learning from videos.

Introducing PhysicTran38K

To train a physics-aware programming model, the authors created a new dataset called PhysicTran38K. It contains about 38,000 pairs of video instructions focusing mainly on body transformation. The dataset covers five major domains:

• Mechanics
• Optical
• Biological
• Important
• What’s hot

In all these domains, it defines 16 subdomains and 46 types of change. Examples include:

• Bright light
• Refraction
• Flexibility
• The cold
• It melts
• Germination
• Strength
• Roll up

Each video captures the full transition from the initial state to the final state, including the steps in between. The construction process is carefully designed and sorted:

• Videos are produced using commands that clearly describe the initial state, start of event, transition, and end state.
• Camera movement filters so that pixel changes reflect physical evolution rather than perspective changes.
• Concrete principles are automatically verified to ensure consistency.
• Only changes that pass this test are saved.

This results in high-quality monitoring of real exchange power studies.

How Does PhysicsEdit Work?

PhysicEdit builds upon Qwen-Image-Edit, a broadcast-based editing core. Incorporating physics, it introduces a two-part approach to thinking:

1. Physically based thinking
2. Clear visual thinking

These two streams complement each other and address different aspects of physical reality.

Dual Thinking: Antecedents of Cognitive Change

Body Based Counseling

PhysicEdit uses the Qwen2.5-VL-7B frozen model to generate a schematic representation before image generation begins.

Given a source image and an instruction, it produces:

• Natural laws involved
• Obstacles to be respected
• A description of how the change should take place

This line of thinking becomes part of the context for setting up a distribution model. Ensures that planning respects cause and domain knowledge.

The mental model remains frozen during training, which helps retain its familiarity.

Implicit Visual Thinking

Thinking about text alone cannot capture visual effects with characters like these:

• Subtle deformation
• Change of state during melting
• Diffusion of light

To handle this, PhysicsEdit introduces readable transformation queries.

These queries are trained using intermediate frames from the PhysicTran38K videos. Two coders guard them:

• DINOv2 properties of structural information
• VAE features texture level details

During training, the model matches transformation queries with visual features extracted from intermediate regions. During inference, no intermediate frames are available. Instead, the questions of the learned transformation act as the driving forces of the transformation, guiding the model to the observable results.

Why is Video Important for Learning Physics?

With image-only monitoring, the model only sees the initial and final states. With video surveillance, it sees how the scene changes step by step. This additional information hinders the learning process. It teaches the model not just what the result should look like, but how it should evolve over time. PhysicsEdit compresses this dynamic information into discrete representations so that editing is always efficient and in one image during projection.

Results on PICABench and KRISBench

PhysicsEdit has been tested in two benchmarks:

PICABench results

PICABench focuses on practical behavior, including optics, mechanics, and state transitions. Compared to its backbone model, PhysicEdit improves the actual physical condition by about 5.9%. The biggest benefits come from categories that require less energy, including:

• Light source effects
• Flexibility
• The reason
• Refraction

KRISBench results

In KRISBench, which tests knowledge-based editing, PhysicEdit improves overall performance by about 10.1%. The improvement is mainly seen in:

• A temporary vision
• Natural science thinking

These results suggest that programming modeling as a change of state improves both visual fidelity and physics-related reasoning.

Why Is This Important For AI Programs?

As generative models are increasingly integrated into creative tools, augmented reality systems, and multimodal agents, virtualization becomes increasingly important. Irrelevant lighting, unreasonable deterioration, or causal breakdowns can reduce reliability and trust.

PhysicsEdit shows that:

• Physics can be learned effectively from video data
• The front of the revolution can be broken down into hidden collective representations
• Textual reasoning and visual supervision can work together

This represents a logical step towards more consistent production models.

Our Top Articles About Photo Editing:

The conclusion

Most planning models treat planning as a problem of constant change. PhysicsEdit reproduces it as a physical deformation problem. By combining video-based monitoring, physical-based reasoning, and learned best practices, it produces planning that is not only mathematically correct but also physically sound. The dataset, code, and testbeds are open source, making them accessible to researchers and developers who want to build realistic programming systems. As productive AI continues to evolve, incorporating physical consistency may move from being a new research topic to a general need.

Note: The source of all images and information on the blog is this research paper.