Saccade 👁️⚡

Don't ask “did the pixels change?” — ask “did anything change that ego-motion cannot explain?”

Benchmark — Full vs TemporalSim vs Saccade

τ — surprisal threshold

energy budget

motion compensation (pseudo-IMU)

method	re-encoded	fidelity	energy	power	hrs @10kJ	budget viol.

Per-frame encoder workload

Shaded bands are motion regimes. Saccade stays low exactly where TemporalSim cannot.

Which patches get re-encoded

frame

TemporalSim —

Saccade —

red = patches re-encoded on this frame

Why it works — the two competing decision signals

same-position residual → re-encode

motion-compensated residual → re-encode

Honest notes

• The stream is deterministic synthetic ego-motion, not natural video — a controlled apparatus so the steady-motion claim is exactly reproducible.
• Energy is an analytic FLOP model (FLOPs × 1 pJ/FLOP), not a power meter. Relative comparisons and the budget guarantee are sound; absolute Joules depend on that coefficient.
• Results use the dependency-free SyntheticBackbone, where each patch embedding is genuinely independent so subset encoding realizes real savings. On a real ViT (DINOv2-small) the same machinery runs but degrades faster at low budgets — see the repo.
• This page is a static Space: every configuration was precomputed offline with the same code the repo's 30 tests exercise, so there is no backend to wait for.