pr0metheus: grep for video
pr0 grep, fast mode, the CLI default, on CPU) to the queries below, over three public-domain films (2 h 49 min).
This free static page cannot run the models, so typing new queries needs the live version: python space/app.py in the repo, or a Gradio Space (Hugging Face now requires PRO for those).pr0metheus is grep for video. Type what you are looking for, get exact time ranges back, each marked confirmed or only closest. It searches what is on screen, what is said, and text that appears in the picture, and it understands "no" and "right after". It runs on a laptop with no paid services.
What this demo shows
Indexes were built with the product's pr0 index --full defaults on a laptop GPU, from the Internet Archive files. Each answer is the product's search (pr0 grep, fast mode) run on CPU: the query is parsed, encoded with the index's text encoder, searched against the picture, speech and on-screen-text indexes, fused, and its edges refined. The vision-model check (--full / --selective) is off, so a result is confirmed only by text evidence: a quoted line heard or read.
Pipeline components and licences
| Part | Model or library | Licence | In this demo |
|---|---|---|---|
| Picture search | google/siglip2-base-patch16-224 (SigLIP 2), the default encoder | Apache-2.0 | Yes: frames and queries |
| Optional picture encoder | timm/PE-Core-B-16 (PE-B, --encoder pe-b) | Apache-2.0 | No |
| Speech index | faster-whisper with Systran/faster-whisper-base.en | MIT (library and weights) | Yes, at index time |
| On-screen text index | RapidOCR (rapidocr-onnxruntime, PP-OCR models) | Apache-2.0 | Yes, at index time |
| Text matching | sentence-transformers/all-MiniLM-L6-v2 | Apache-2.0 | Yes |
| Vision-model check | Qwen/Qwen3-VL-2B-Instruct, 4-bit | Apache-2.0 | No: only with --full / --selective |
Benchmark heads are non-commercial and not part of the product. Heads trained for benchmarks (QVHighlights, TVR) and research setups such as PE-L fusion or Qwen3-VL-8B re-ranking are benchmark-only. Benchmark data never enters product training.
Films
Public-domain status is for the United States. Video here is a 360p re-encode for the player, on the same timeline as the files that were indexed.
Indexing time
All three films (2 h 49 min of video) were indexed together in one run of pr0 index --full with the shipped defaults, on a laptop with an NVIDIA RTX 3050 (4 GB):
| Stage | Ready after | What it covers |
|---|---|---|
| Visual search | 116 s | 10,158 frames, 87.7 frames/s |
| On-screen text | 863 s | 465 text segments |
| Speech (full index done) | 979 s (16.3 min) | 3,379 speech segments |
That is about 41 s per hour of video until visual search is ready, and about 5.8 min per hour for the full index. The films were indexed as one batch, so separate per-film times were not recorded. Speech took the longest because His Girl Friday is very dialogue-dense. For comparison, on a Kaggle NVIDIA T4 one hour of video is visually searchable in about 53 s and fully indexed in about 6 min.