Hi Thanks for sharing great work!
Neoverse supports a variety of applications, and I’d like to better understand how each of them works.
For example, in the case of bullet time, is it implemented by adjusting a user-defined trajectory into a bullet-time-style camera path? Is it also possible to try this feature in the demo?
Also, for counterfactual simulation, I noticed that objects like elephants or dinosaurs appear in dashcam videos. Are these generated by conditioning a diffusion model on text prompts, or are long-tail objects explicitly added into the 4D Gaussian representation?
Finally, when training the diffusion model, what kind of captions are used? When I ran the demo (app.py), the default prompt was:
"A smooth video with complete scene content. Inpaint any missing regions or margins naturally to match the surrounding scene."
Is this the same type of caption used during training, or is it just a default inference-time prompt?
Thanks.
Hi Thanks for sharing great work!
Neoverse supports a variety of applications, and I’d like to better understand how each of them works.
For example, in the case of bullet time, is it implemented by adjusting a user-defined trajectory into a bullet-time-style camera path? Is it also possible to try this feature in the demo?
Also, for counterfactual simulation, I noticed that objects like elephants or dinosaurs appear in dashcam videos. Are these generated by conditioning a diffusion model on text prompts, or are long-tail objects explicitly added into the 4D Gaussian representation?
Finally, when training the diffusion model, what kind of captions are used? When I ran the demo (app.py), the default prompt was:
"A smooth video with complete scene content. Inpaint any missing regions or margins naturally to match the surrounding scene."
Is this the same type of caption used during training, or is it just a default inference-time prompt?
Thanks.