Search video filenames, tags, imported captions, scene colors, and visually similar frames — all on-device without a model. Optional CLIP adds sampled-frame text/image matching only after you enable it.
🔒 Standard indexing and search run locally, with no AI download. Optional CLIP loads only after you select it and approve a model download. Your footage is not uploaded.
03 Search video moments
Not indexed
Standard metadata / captions (no AI)
Try:
Add a video or try the demo, then Build Standard index. AI stays off unless you enable it.
Videos0
Indexed frames0
Scenes estimated0
Search methodStandard / No AI
Moment player
No clip selected
▶Your scenes, instantly accessibleChoose a result to jump to its timestamp.
—
Matching moments
0 results
⌕
Your best matches will appear here
Build an index and describe a visual moment to search across uploaded videos.
Find visual matches, scene colors and your own captions without AI
PixelOkay Video Search Engine first builds a local Standard index from video thumbnails, RGB histograms, spatial color and edge features. It can search filenames, user tags, notes, imported SRT/VTT captions, dominant colors and visually similar frames. Optional CLIP can compare sampled frames to descriptions if explicitly enabled and downloaded.
No model download is needed for Standard search. If the user opts into AI, static runtime/model files may be requested from a third-party host; inference stays local. Neither mode reliably observes events between sampled frames, and Standard metadata search cannot invent captions or recognize objects. For accurate large-scale forensic indexing, use professional tooling and verify all returned clips manually.
Using Video Search Engine
Search sampled moments across local video clips with text prompts, reference stills or an already indexed frame. Build an index first, then open candidate results at their source time to confirm what is actually visible.
Add up to five videos or choose the bundled demo, set the sampling interval and build the visual index.
Search with a short visual description or reference image; you can also compare a new frame against indexed frames.
Review the ranked frame candidates and source timestamps, browse nearby scenes, then export selected markers, stills, a contact sheet or the index.
Example: After indexing three short interview clips, search for “person entering a room,” open promising sampled frames in the source player to confirm the moment, and export only the verified timestamps.
Limits and privacy: The index is limited to five videos and 360 sampled frames; sparse sampling can miss brief events. Similarity is a ranking score, not a probability or temporal action detector. Choosing CLIP indexing loads runtime/model files from external hosts and may disclose the visitor IP. Saved indexes and thumbnails contain sampled frame content; sampled fingerprints are not full-file checksums. A completed CLIP index, inference and useful retrieval/export have not yet passed runtime acceptance.
Frequently asked questions
Does the search inspect every frame?
No. It embeds sampled frames, up to 360 across the library. A short moment between samples may not appear, so inspect the source player and adjust sampling when practical.
Does a high similarity score prove that the described event happened?
No. The score ranks visual similarity for the selected text or image query. It is not a calibrated probability, and the tool does not establish what happened between sampled frames.
