01
Use a cleared corpus
Publish the clip list, rights status, duration, codec, frame rate, resolution, and audio properties. Keep the same corpus across tools.
Focus Research · Revised 29 August 2026
Measure video search without turning a pilot into a promise.
A disclosed one-run Focus indexing pilot on Apple M1 Pro, its limitations, machine-readable summary, and versioned open tables for evaluating indexing and retrieval honestly.
CC BY 4.01 run disclosedNo product comparison
Short answer
What does this pilot actually show?
On 5 January 2026, a MacBook Pro with an M1 Pro and 32 GB of memory processed a 97.685-second H.264 720p clip at a 0.76 real-time factor. That is about 46 seconds of processing per source minute. This result covers one local-ingestion run.
0.76×
real-time factor
M1 Pro
32 GB memory
1
pilot run
Pilot method
The recorded facts
Missing details are listed as limitations instead of being reconstructed after the fact.
| Run date | 5 January 2026 |
|---|---|
| Device | MacBookPro18,1; Apple M1 Pro; 10-core CPU; 32 GB memory |
| Source clip | 97.685s; H.264; 1280x720; 24 fps; AAC |
| Pipeline | Local ingestion including transcription, visual analysis, scene detection, embeddings, and indexing |
| Runs | 1 |
| Measured result | 0.76 real-time factor |
The data file and method are licensed CC BY 4.0. The source clip is not included in the download.
One observation, without extrapolation
Source duration
97.685 s
Calculated processing time
≈ 74 s
Processing time is derived from source duration × 0.76; it is not a second measured run.
Limitations
What this result does not prove
These limits prevent the pilot from being used as a product ranking or universal estimate.
This is one run on one short clip. It is not a population estimate or a hardware comparison.
The internal source clip cannot be redistributed, so the original run cannot be reproduced bit-for-bit.
The original output did not record app version, macOS version, elapsed seconds, thermals, or power state.
The run measured ingestion time. It did not measure search precision, recall, user task time, or another product.
Open protocol v1
How to produce a defensible benchmark
The v1 bundle separates corpus, runs, queries, relevance judgments, and ranked results.
02
Record the machine and build
Capture app version, commit, operating system, device model, chip, memory, power state, and cold or warm cache state.
03
Repeat timing runs
Run at least three cold and three warm trials per clip. Report every observation, median, range, failures, and real-time factor.
04
Judge retrieval separately
Pre-register speech, visual, person, and scene queries. Label relevant moments before testing and report precision@5 and recall with human review.
Corrections or independent reproduction: [email protected]
Protocol v1 files
README · corpus.csv · runs.csv · queries.csv · qrels.csv · results.csv · CITATION.cff · CHANGELOG.md · SHA256SUMS.txt
Test the workflow on your own footage
The free trial indexes up to 30 videos. Results depend on your hardware, selected models, and media.