Batch auto-annotation and review

Choose exactly what the job may touch

• Whole dataset — normally restricted to files that have no accepted annotations yet.
• Current selection — the files checked in the Media grid.
• Current filter — every file matching the active saved or temporary filter, not only the loaded page.
• Semantic search results — the matched still images from a web-or-desktop search.

Prediction model accepts native videos as well as images. Text prompt, Visual prompt and external-assistant batches accept still images; extract frames to use these engines with video content. If annotated images are explicitly included, accepted work remains unchanged and the job adds separate drafts beside it.

Prediction sources

The dialog computes availability from the project, device and active model state. An unavailable source remains disabled with a reason rather than falling back to another model. Score scales differ by engine, so the threshold uses an engine-specific default and is remembered per project. When the source will run inference, the dialog names the actual model beside the result type — for example MobileSAM + CLIP or SAM 2.1 Small + SigLIP 2 — rather than implying that every installed SAM participates. Classification Text Prompt reuses stored Semantic search scores and deliberately does not claim that the currently selected model produced them.

Batch video annotation with Prediction model

Include videos in the dataset, selection or filter and choose Prediction model. Video sampling offers every frame, every 5 frames (default), every 10 or every 15. The dialog shows the number of model inputs and always includes the final frame. One pinned model runs sequentially through the samples; progress reports both completed files and sampled frames.

Every frame runs inference directly. Larger steps reuse the annotator’s existing interpolation: matching same-label boxes interpolate linearly, while polygons and classification results hold their previous sample. This is not tracking with stable object identities; sparse sampling can miss brief appearances or fast motion. The interactive annotator’s automatic sampling is approximately five predictions per second, not a fixed five-frame step.

A completed video is saved atomically as Pending AI Review. Open first item enters the video annotator; scrub to inspect the generated frames, edit and save individual frames, then use Accept video or Reject video for the remaining drafts. The gallery’s existing bulk Accept/Reject actions also work. Saved manual frame edits survive rejection. Drafts stay out of training exports, dataset versions and accepted annotation statistics.

Cancel or a frame timeout pauses a video job. Finished files remain reviewable; Resume restarts the unfinished video using its saved sampling step and pinned model. A partial video is not saved. Full project archives preserve pending review state; ground-truth dataset exports exclude it.

See what the batch is doing

The selected AI Assistant is the fourth card, after Visual prompt. Starting sends eligible images and project labels to the selected provider, OpenAI or Anthropic: up to two concurrent API requests, or one image at a time through desktop Codex or Claude Code. It supports boxes, segmentation polygons and image classification. Create labels in AI Assistant before the batch. Results use the same Pending AI Review workflow, with no fabricated confidence scores or score-threshold acceptance. See the AI Assistant guide for connection setup and billing.

While the job runs, the dialog keeps the completed-file count separate from the work inside the current image. It names the current file and the actual stage — preparing a model session, loading or encoding an image, evaluating prompts, generating or scoring proposals, optional refinement, or saving results — with counters when the model reports them. Job, current-image and current-stage timers advance from service measurements rather than from the moment the dialog happened to open.

The finished view reports active job time, average inference time per completed file, annotations actually saved, images with results, images where the model found nothing, remaining work after a stop and any per-image failures. The average includes model preparation and successful empty-result calls, while skipped images and failed inference calls stay out of its denominator. A resumed job whose older checkpoints lack complete measurements is labelled as recorded work rather than presented as whole-run statistics.

Reviewing Pending AI work

1. Use the Pending AI Review filter or the Review count in the Media toolbar.
2. When a run finishes, open its first still-pending item directly in the annotator, or open the review gallery for the batch.
3. Sort by AI score to inspect the strongest drafts first, or reverse the order to hunt for likely mistakes. A Text Prompt Match score is similarity on that method’s scale, not a probability.
4. Accept all, Reject all or accept everything at or above a selected score from the review bar.
5. Open an item for object-level correction. Accept saves what remains as ordinary annotations; Reject removes only that job’s drafts.

Safety and interruption

• Local engines run sequentially on your device. The optional ChatGPT or Claude engine sends images online, with up to two API requests at a time or one desktop Codex/Claude Code request.
• External-assistant cancellation, authentication errors, quota errors and rate limits pause the job; completed drafts remain reviewable. Reopen Auto-annotate to resume with the original prompt, model, connection and label schema. Requests are not automatically retried.
• Cancelling stops future inference and keeps already written results in the review queue.
• An image edited by a person while a bulk action is running is skipped and reported.
• Rejecting a run removes its drafts, never accepted annotations that existed before it.

iPhone and iPad do not offer Semantic search, Text prompt or custom-model setup. Batch work there is limited to a prediction source the mobile build can actually prepare, such as Visual prompt from existing annotations. External-assistant batch annotation is unavailable on native iOS and Android.