Problem
Fans see a match from the broadcast camera, never from the perspective of the player making the decisions.
Solution
Scene understanding through video models identifies the subject; monocular depth estimation builds a coarse 3D scene and a rough first-person sketch video; an image-to-video diffusion model fine-tuned with LoRA on paired third- and first-person clips renders the final first-person footage on cloud GPUs.
Demo moment
Team-described: pick a player in a soccer clip and watch the play from their eyes.
Why it stood out
- organizer statedTook second place overall among 160 submissions and two sponsor prizes.Source ↗
- team statedFine-tuned a 14B video diffusion model on paired viewpoints during the event, with depth-based priors to keep camera motion plausible.Source ↗
- curator inferenceSeeing a famous player's point of view is an instant hook for any audience.
Lessons for your next hackathon
- Training a model in a weekend is credible when the output is something judges can watch and compare.