08 / Video prediction
Video Net
Compare real model predictions with the motion they try to continue.
Saved model output · two authored synthetic probes
08 / Actual model predictions
How long does motion survive?
The model observes four frames of an authored animation, then predicts 48 frames using its own previous output. Scrub to compare each prediction with the source.
Loading model output…
Error across the rollout
Two synthetic probes outside the model's UCF101 training domain. Predictions come from the saved 30,000-step checkpoint; playback uses precomputed frames.
The project
Predicting motion, then trusting it
A flow-and-residual U-Net takes four RGB frames. It warps the latest frame with predicted optical flow, adds a learned residual, and returns the next frame. During a rollout, that output becomes part of the next input.
Observations & lessons
What this probe shows
The saved model initially preserves the synthetic scene, then motion and detail drift as it consumes its own predictions. The error plot measures that drift against a known continuation. These authored clips are a new diagnostic outside its human-action training domain.