Key Features

Controllable car crash video generation
Conditions on bounding boxes, crash types, and initial image frame
Classifier-free guidance with independently tunable scales
Fine-grained control at inference time
State-of-the-art performance on quantitative video quality metrics
State-of-the-art performance on qualitative measurements of physical realism and video quality
Ability to generate counterfactual scenarios
Realistic car crash generation

Ctrl-Crash is designed to generate realistic car crashes, which is a challenging task due to the scarcity of accident events in most driving datasets. The model uses a combination of bounding boxes, crash types, and an initial image frame to generate a realistic crash scenario. The model can also generate counterfactual scenarios, where small changes in the input can lead to significantly different outcomes. This feature makes Ctrl-Crash a valuable tool for applications such as traffic safety and accident reconstruction.


Ctrl-Crash has been evaluated on various metrics, including FVD and JEDi, and has achieved state-of-the-art performance. The model has also been compared to other methods, such as Cosmos, Sora, AVD2, DrivingGen, and Ctrl-V, and has outperformed them in terms of video quality and physical realism. The model's ability to generate realistic car crashes and counterfactual scenarios makes it a valuable tool for various applications, including traffic safety, accident reconstruction, and autonomous vehicle development.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!