LAION · AUDIO + VIDEO · CAPTION QUALITY
Original captions, cross-model revisions, and evidence
SALMONN and MiniCPM each describe 50 AI-labelled and 50 comparison videos. Then each model revises the other model’s caption while inspecting the video and soundtrack again. Compare all four variants, their Gemini 3.8 Flash ratings, and their measured processing times.
Every analysis covers the first 30 seconds or less. Generated and revised captions can contain mistakes; Gemini ratings are automated judgments, not human ground truth. Retrieved captions describe other source videos. The comparison cohort has dataset AI flag zero and has not been independently verified as human-created.
Completeness and factuality, scored separately from 0 to 5
| Caption variant | Cohort | Rated / available | Mean completeness / 5 | Mean factuality / 5 |
|---|
Did revision improve each original caption?
Paired changes compare the revision with its own original on the same videos. A positive delta indicates a higher automated score; worse revisions are retained.
| Revision pipeline | Cohort | Paired videos | Completeness Δ | Improved / same / worse | Factuality Δ | Improved / same / worse |
|---|
Measured end-to-end caption processing
| Caption pipeline | Videos | Original generation mean | Revision mean | End-to-end mean | End-to-end median | End-to-end p95 |
|---|
Model initialization and Gemini API time
Models, revision instructions, evaluation rubric and retrieval methods
Revision instruction
Gemini scoring rubric
Previously measured PEAV embedding and CPU retrieval throughput
| Model / operation | Measured sample | Time / throughput | Scope |
|---|