Monitor and control the Celery-based import queue system.
Note: Failed/skipped tasks older than 7 days are auto-cleaned daily.
| File | Type | Status | Error/Reason | Time |
|---|---|---|---|---|
Configure filters to skip certain images during import.
Permanently delete all images associated with a specific tag.
Scan and remove images that don't meet filter criteria.
Images with more than this % transparent pixels will be flagged.
Scan for visually similar images using perceptual hash comparison. The first image in each group has the highest resolution and will be kept.
Lower values find only very similar images. Higher values find more potential duplicates but may include false positives.
Background operations and worker management.
Workers: - | Active Tasks: -
Run inference on all images missing predictions or embeddings for the current model versions.
Resumable and safe to re-run. Backfill enqueues work on the ml queue; monitor progress in the ml-worker logs.
Recompute centroids for all eligible tags (character, fandom, and general/topic) that have at least the minimum number of reference images. Run this after the initial backfill, or any time you've applied many tags manually and want the suggestions to catch up.
For every character tag with a fandom assigned, attach the fandom tag to every image already tagged with that character. Additive only — never removes fandom attachments. Safe to re-run. Backfills data migrated from pre-refactor tag names that embedded the "(Fandom)" suffix, since the migration set tag.fandom_id but did not retroactively attach the fandom row to image_tags.
Walks every image and runs a structural check (PIL for images,
ffprobe for video, zipfile for archives). Flagged rows are excluded
from random/showcase, ML retag, and the suggestion modal until a
fresh copy lands in /import/<artist>/<filename>
— at which point the import pipeline supersedes the corrupt row,
preserving its tags. The default mode only checks rows whose status
is still unknown (post-migration default + freshly
imported); use "force re-verify all" after disk events to recheck
everything. Flagged rows are listed at
/api/integrity/failed.
Walks every image and applies general-category WD14 predictions whose confidence is at or above the auto-accept threshold. Same side-effects as opening each modal individually (creates the kind='user' tag if needed, attaches it to the image, logs a SuggestionFeedback rejection-protected record). Idempotent — safe to re-run. Useful after a fresh ML backfill so high-confidence tags land without needing to view each image. Skips images flagged as integrity-corrupt.
When the modal loads suggestions for an image, any general-category tag with confidence at or above this threshold is auto-applied and shown in a reviewable "Auto-accepted" block at the top of the suggestions section. Lower values accept more tags (riskier); values above 1.0 effectively disable the feature. Default is 0.95.
One canonical tag name per line. Matches are exact against the
post-canonicalization WD14 output (e.g. 1girl,
breasts). Blocklisted names never appear in the
modal's auto-suggestions — either WD14 or embedding-similarity
sourced. Adding a name also enqueues a cleanup task that removes
the tag from every image that has it attached (kind='user' only;
deliberate character / artist / fandom tags sharing the same name
are left alone). You can also add entries inline by clicking the
⊘ button on any suggestion chip.