08 / Video prediction

Video Net

Frame prediction and the error that builds during rollout.

Method overview · footage under review

Explore this project

Video Net takes a short stack of frames, estimates motion with a flow-and-residual network, then feeds its own predictions back into the next step.

The study follows what happens when a one-step prediction becomes input for another. Small errors in motion and detail can accumulate across a rollout.

Experiment boundary

What the comparison will show

01 / ObserveA short stack of source frames
02 / PredictFlow plus residual next-frame estimate
03 / Roll outFeed each prediction back into the model
04 / CompareInspect drift against the true frames

A frame comparison will follow when the research footage is cleared for this page.

Reading a rollout

01

Observe

The model receives a short stack of source frames and predicts motion plus residual detail.

02

Predict

Its output becomes an input at the next step, so errors can accumulate across time.

03

Media

The local study has saved rollouts. The source footage is being checked before any clip is placed on this public page.