Reproduction tutorial · September 2, 2026
Adaptive video sampling versus static frame rate
These tutorials test the mechanism behind each paper at a practical scale. They do not reproduce frontier-scale headline results unless explicitly stated.
Source of this tutorial
This tutorial tests the mechanism behind reviewed development 02 in the immutable briefing of September 2, 2026.
Scope and limits
- Reproduction level
- partial mechanism reproduction
- Claim under test
- a scan-then-zoom sampling policy can answer video questions with materially fewer frames than uniform sampling without a material accuracy loss.
- Excluded claim
- this tutorial does not reproduce Google's reported 88 percent token reduction, 66 percent cost reduction, or 7 percent accuracy gain.
- Reader, effort, and cost
- intermediate Python; 5 to 8 hours; local CPU with optional modest GPU; keep any model API budget below $20.
- Prerequisites
- Python 3.11, ffmpeg, and a video question-answering model or scripted scorer. Record resolved versions and licenses.
Canonical resources
- Google, introducing agentic video in Gemini: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-agentic-video-in-gemini/ (opens in a new tab)
- Corpus notice: Use 20 openly licensed annotated videos with published temporal annotations. Record each license at execution time.
Project plan
| Step | Stage | Effort |
|---|---|---|
| 01 | Assemble the corpus | 60 minutes |
| 02 | Static baseline | 60 minutes |
| 03 | Adaptive condition | 90 minutes |
| 04 | Seeds and ordering | 45 minutes |
| 05 | Evaluate | 60 minutes |
| 06 | Cleanup | 15 minutes |
01Assemble the corpus(60 minutes)
Select 20 openly licensed videos with temporal annotations. Freeze the query set, the total frame budget, and the scoring rubric before any run.
02Static baseline(60 minutes)
Sample uniformly at 1 frame per second. Log accuracy, temporal intersection over union, frames used, tokens, latency, and miss rate for brief events.
03Adaptive condition(90 minutes)
Scan at 0.1 frames per second, select candidate windows, then resample those windows at 4 frames per second. Keep the same query set, frame budget, and scoring.
04Seeds and ordering(45 minutes)
Run five query-order seeds per condition so ordering effects are visible in the intervals.
05Evaluate(60 minutes)
Report accuracy, temporal intersection over union, frames, tokens, latency, and brief-event miss rate with bootstrap intervals. Success requires at least 50 percent fewer frames without a material accuracy loss.
06Cleanup(15 minutes)
Delete extracted frames, retain the query set, seeds, per-run metrics, and aggregate tables.
Controls, failure modes, and safety
Controlled variables
corpus, query set, total frame budget, scoring rubric, model, seeds, and latency measurement method.
Failure modes
budget mismatch between conditions, annotation leakage into window selection, brief-event blindness, scorer variance, and latency measured without decode cost.
Safety
use openly licensed material only; do not process private or personally identifying video.