-
Notifications
You must be signed in to change notification settings - Fork 0
Troubleshooting
No frames from it were described. Almost always one of:
- No vision model. Settings says which model fills each role, and what happens for the ones nothing fills. This is the role that matters.
- The provider is off. A model can be configured and its provider disabled; the Settings summary marks that case.
- The scene is short and dark. Blank frames and fades are screened out before a model ever sees them, because a model asked to describe black will invent something rather than say so. A scene made entirely of them has nothing to describe.
MAN 7 sits. WOMAN 2 stands. means no text model. Without one, Capra assembles the
script from what it measured rather than having a model write it — richer than it used to
be, and still plainly not prose.
Add a text model in Settings, then Rewrite all on the Script page. Locked lines and anything you edited are left alone.
Some of this is expected and some is a signal.
- Extras. On a real film most of the cast never speaks and appears in one scene. Those are marked as extras, numbered, and never asked about. If they are cluttering the page, that is what the search box and the named-first ordering are for.
- Duplicates. Merge them: drag one card onto another, or use the Questions page, where the film is playing beside the question.
- Three frames on one card and none of them the same person. Open the frame and press Not this person — that takes the whole appearance off the character and gives it its own.
Capra learns from those answers. After three, it stops offering pairs no more alike than ones you have already refused, and says so on the first question.
A cast card showing "no frames yet" means no frame carrying that subject was ever extracted. For a character with no detection at all, Capra falls back to the stretch of film they were seen in — so this should now be rare. If a whole page is empty, the keyframe extraction stage has not run or has failed; the analysis log says which.
Measured on a 26-minute film: 46 minutes of wall clock, two thirds of it vision calls.
The lever with by far the biggest effect is a cloud vision provider — those calls are latency rather than compute, so six at a time overlap almost perfectly and a forty-minute pass becomes a few minutes.
Locally, the options are fewer frames or a smaller model. Both cost something real, which is why neither is the default.
Face embedding is the next largest at about a fifth of the time, and it runs on the CPU when onnxruntime cannot reach the GPU. It can be turned off in the advanced settings: the cost is a worse cast list rather than a broken run.
The proxy is missing or was interrupted. Capra writes it to a temporary name and renames it only when ffmpeg finishes, so a half-written file is never adopted — but a proxy job that never ran leaves nothing to play. Check the analysis log for the proxy stage.
The log says answers were cut short by the model's token limit. A frame with several
people fills the observation schema with a lot of JSON, and a truncated answer parses as
nothing at all. Use a model with more headroom, or one that honours a JSON schema
properly — Capra records which yours does when it is added.
Pick the language before starting, not after. Capra writes the screenplay directly in the language you choose rather than translating afterwards.
If characters appear as black boxes anywhere, that is a font fault and a bug — please report it with a screenshot. The bundled faces cover Turkish, Polish, Hungarian, Czech and the rest of Latin Extended-A.