RealWonder turns single images into real-time physics simulations

RealWonder is a newly open-sourced video generation system that lets you apply virtual forces directly onto a single static image and watch objects move in real time.

Most video models today have zero clue how physical forces actually work in 3D space. RealWonder fixes this by letting you drag an arrow on your screen to push, blow, or drag objects interactively, seeing the physical reaction immediately.

What you can do with it

You feed the system a single picture, and it turns into an interactive playground.

Drop a force arrow onto an object floating in water, and it flows along the simulated current. Direct a gust of wind at a hanging shirt, and the fabric billows in that direction. You can even apply multiple forces simultaneously—pushing from the left while pulling from the right—and it resolves the combined dynamics smoothly.

The framework supports rigid objects, deformable bodies like cloth, fluids, and granular materials like sand. It also handles simulated robotic manipulations and camera path controls.

In terms of raw speed, RealWonder hits 13.2 frames per second at 480x832 resolution on a single Nvidia H200 GPU. While 13 FPS is not esports-level smooth, it firmly crosses the threshold into real-time interactive generation for AI video.

How it works under the hood

Instead of forcing an AI model to guess complex physics from scratch, RealWonder uses a classical physics simulation as a bridge.

The pipeline integrates three key components: single-image 3D reconstruction, a physics simulator, and a distilled video generator. When you apply a force, the simulator first calculates the motion in 3D and converts continuous actions into visual cues like optical flow and RGB renders, which video models naturally understand.

The video generator itself is heavily distilled, requiring just 4 diffusion steps to produce each frame. That dramatic reduction in compute steps is what makes the 13.2 FPS speed possible.

Both the complete codebase and trained model weights are publicly available on the project website and GitHub, along with step-by-step local setup instructions.

Why it matters

We think RealWonder represents a smart pivot from passive video generation to actionable physical simulators.

Instead of blindly prompting and praying the visual physics look right, robotics researchers and game developers now have a proof-of-concept pipeline for real-time interaction from static images. If you work in robot learning, VR, or physical scene generation, this is an immediate clone-and-test repo.

The catch, of course, is hardware. Pulling 13 FPS requires a high-end H200, putting it out of reach for consumer rigs right now. But as an architectural pattern—bridging 3D physics engines with 4-step distilled diffusion—it shows exactly how real-time generative worlds will be built.