Thanks to my collaborators at @mila-quebec.bsky.social for making this project possible: Ge Ya Luo, Luis Lara (@ludolara.bsky.social), Florian Golemo (@shoddyrobots.bsky.social), Derek Nowrouzezahrai, Liam Paull, Alexia Jolicoeur-Martineau (@alexiajm.bsky.social), and Chrisopher Pal.

The generated clips are 1) Left 2) Right 3) Left

Details such as text and bluriness may give it away. However, Ctrl-Crash produces realistic physical details such as cars shaking and items on the dash moving on impact. We trust that better visuals can be achieved through scaling.

3)

2)

🕹️ Can you guess which video is a real dashcam clip and which is a generated by Ctrl-Crash using only the initial ground-truth image, 2D bounding-boxes for the vehicle motion, and the desired crash type?
1)

Our project is fully open-source:
📄 Paper: arxiv.org/abs/2506.00227
🌐 Website: anthonygosselin.github.io/Ctrl-Crash-P...
🔗 GitHub: github.com/AnthonyGosse...
🤗 HuggingFace: huggingface.co/AnthonyGosse...

🚗💥Introducing Ctrl-Crash: controllable video generation for autonomous driving! SOTA models struggle to generate physically realistic car crashes. We propose an image2video diffusion model with bounding box and crash type control.
Website: anthonygosselin.github.io/Ctrl-Crash-P...
🧵->