fc1e8b9c33
Videos were never part of the automated pipeline before -- only reachable one at a time via a manual "Tag Now" button, and even that just grabbed one static frame and ran it through the photo model. Now videos are first-class: - New extract_video_frames_b64(): samples up to 6 frames spread across the clip's duration and passes them all to the video model in one call, so it sees actual motion/progression instead of one snapshot. Verified live: two different real test videos got distinct, content-aware descriptions that correctly named what was actually happening in each, not generic placeholders. - ollama_generate() now accepts a list of images (photos still pass a single one, unchanged) so the same call path serves both. - process_loop merges S.video_files into the same pending queue as photos, routes videos to a separate configurable video model (default minicpm-v4.6:1b, a small dedicated vision model) and prompt, skipping the photo/screenshot router entirely. redo_single (the lightbox "AI Redo" button) updated the same way for consistency. - S.total_images is now set to the actual combined pending count for the run so the progress bar/ETA reflect videos too, not just photos. - FOUND WHILE TESTING: write_metadata's video branch used "-Keywords" for the category/photon-tagged marker, which is a silent no-op on QuickTime/.mov files (exiftool has no mapping for it there) -- confirmed by direct testing. Every video "tagged" before this would have gotten a description but never an actual category keyword embedded. Switched to XMP-dc:Subject (the same tag family used for photos, which QuickTime containers do support via an embedded XMP packet) -- verified the category now lands and stays idempotent across repeat writes. - New frontend controls: video model dropdown and a "tag videos too" toggle (on by default) in Engine Configurations. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
104 KiB
104 KiB