Model card: Shotlore keywords, captions and transcripts
Shotlore doesn't train or host a model. It sends a small preview of each still or clip to a model you choose, with your instructions, and shows you what comes back for approval. This card says which models, what goes in, what comes out, how it fails, and where the person is.
What it is for
Suggesting keywords, a caption and a short headline for each still or clip in a job, so a shoot arrives in an archive or a library described and searchable; optionally, writing out what is said in a clip. Nothing is published by Shotlore: every suggestion is reviewed, edited or rejected by the person running the job before it is written into a file or an export.
Models
| Route | Model | Provider and terms | Used for |
|---|---|---|---|
| Gemini (default for new jobs) | gemini-3.5-flash-lite | Google, Gemini API, under the production's own key and Google's Gemini API terms (paid or free tier: see Security). | Keywords, captions, headlines; transcripts; writing tagging instructions from a shoot description. |
| Gemini (earlier jobs) | gemini-2.5-flash | Google, as above. If Google won't serve it to a key, Shotlore switches to 3.5 Flash-Lite and says so. | As above. |
| On this Mac | Apple Vision image classification (the macOS framework, no network) | Apple, on the computer; nothing leaves. | Keywords only; no captions, headlines or transcripts. |
| People shown (optional) | Apple Vision face detection and feature prints, on the computer | Apple, on the computer; nothing leaves. | Grouping faces within one job so you can name them; names are written as Person Shown. |
The model name used for each call is recorded in the job's send log.
Inputs
- A JPEG preview of the still, or frames spread evenly through a clip at a rate you set (1 to 12 a minute of footage, 3 by default; at least 3 and at most 60 a clip), each with the time in the clip it was taken, longest side 1600, 2048 or 3072 px (2048 by default). Never the original file or a full-resolution frame.
- Your tagging instructions: a preset (general or natural history), text written from your shoot description, or text you edited. Plus, for consistency across a job, up to 80 keywords already approved in it and, if the client has one, the client's keyword list.
- Unless the job is confidential: the capture time; the position (exact, or rounded to about 10 km or withheld for a sensitive place); the place name (withheld for a sensitive place); and the one-line speech summary if the clip was transcribed.
- For transcripts: the first 45 minutes of the clip's sound as 32 kbps mono AAC, and your shoot description for context.
Outputs
- A JSON object with
keywords(a list),captionandheadlineand, for a clip,shots(a timed list of framing changes and cutaways, from 0.7.6), constrained by a response schema; temperature 0.2. Keywords are lower-cased except proper names, de-duplicated, and passed through the client's keyword list if there is one. - For transcripts:
transcript(speakers marked, "[inaudible]" rather than guesses) and a one-linesummary. - Google also returns token counts, which the send log records.
The human in the loop
Every suggestion appears on the file's card marked as not yet approved. The person approves, edits or discards it; nothing is written into a file, a sidecar, a report, a list or a manifest until it is approved, and "Save keywords" is a separate step. Transcripts are shown in full and can be corrected. The job records what was approved; a bulk edit can change many at once but is still the person's action.
Known failure modes
- Confident wrong species, places and names. A model may name the wrong antelope, call a village by a nearby town's name, or invent a person's name from a caption-like cue. Treat every proper name as unverified until you check it.
- Place names from the position, not the picture. When a position goes with the preview, the caption may mention the place even if it isn't visible. On confidential jobs no position goes.
- Drift between files. Wording and casing wander across thousands of files; the job's reuse list and "Tidy spellings" reduce it, not remove it.
- Frames that aren't representative. Three frames a minute of a long clip can miss what matters; raise the frames per minute for important clips.
- Transcripts mis-hear names and numbers, mix speakers in crosstalk, and can miss quiet speech; they are a draft, not a record.
- Refusals. Google may decline a frame (safety filters); Shotlore shows the error and moves on.
- Text in the frame. Responses are constrained to the schema, so a sign or slate in shot can't instruct the model into doing something else; it can still bias the caption. (OWASP LLM top-10 prompt-injection review: documented, not independently tested yet.)
- Bias. The models are Google's general-purpose models; their descriptions of people, places and cultures reflect their training. Read captions involving people with that in mind, especially for broadcast.
Data handling
- Transport: TLS to Google. Storage by Google: on the paid tier, not used to improve Google's products; kept 55 days for abuse monitoring, in any country where Google has facilities; see Security for the quotes.
- On the computer: previews in the job's cache folder; suggestions in the job file; the send log in the job file.
- Sensitive places: positions rounded or withheld and names withheld before anything is sent.
Evaluation
Shotlore's own tests cover the request and response handling, the rounding of sensitive positions, the send log and the schema. We have not published a benchmark of keyword or caption accuracy for these models on natural-history or documentary rushes; productions should spot-check a sample of each job, which the Review step is built for. We will add figures here when we have them.
Versions
This card describes Shotlore 0.9.0 (the models and inputs are unchanged since 0.7.4; the shot list output arrived in 0.7.6). Changes to the models, the inputs or the schema are listed in What's new, and this page's date changes with them.