Videos now route through a 10-frame sampling branch (configurable via
VIDEO_ML_FRAMES env) instead of the previous unsupported_format early
return. WD14 predictions are aggregated by max-confidence per (name,
category) across frames so sparse signals aren't diluted; SigLIP
embeddings are mean-pooled for a representative shot. Also generates a
fallback thumbnail when the record is missing one, and removes the
video_filter from backfill so videos get enqueued.
Celery soft/hard limits bumped to 240s/360s to accommodate 10x inference.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
bvandeusen
bc6b0a4a58