> For the complete documentation index, see [llms.txt](https://docs.unitlab.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.unitlab.ai/documentation/annotations/ai-assisted-annotation.md).

# AI-assisted Annotation

Unitlab’s assists accelerate proposal creation; they do not change who is accountable for the label. Use the smallest assist that matches the task, correct every proposal, and measure systematic errors before expanding volume.

### Before you make the change

* Define the class, geometry, output, acceptance rule, and human owner.
* Choose a representative calibration cohort that exposes model weaknesses.
* Know the model or assist version and the route for empty, malformed, low-confidence, or failed output.

![AI-assisted Find Similar results](https://292810646-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FGjVLUz4wthGkGlRKM6rM%2Fuploads%2FmoLNHYRw2eyGq3iAMLTT%2Ffind-similar-results.png?alt=media\&token=dda7d9ce-1786-4420-96ba-4daca8bb2f6b)

*Assisted results should be reviewed as proposals: inspect false positives, missed instances, class mapping, geometry, and source context before acceptance.*

### Understand the product behavior

Unitlab combines local assistance inside the workbench with Model stages inside workflows. Assisted results remain proposals until they are reviewed and accepted through the annotation and quality process.

### Magic Touch

Magic Touch is available in image and video toolbars as an interactive mask-oriented tool. When no compatible class is selected, it opens a Create Class dialog configured for a Mask class, with name, color, and numeric hotkey fields.

The correct production loop is:

1. Select the target mask class.
2. Guide the proposal on a representative object.
3. Inspect boundaries, holes, thin structures, and occlusion.
4. Correct with manual tools.
5. Review the final mask under the same policy as a fully manual label.

### Detect all objects

A class-aware auto-labeling panel contains:

* current class;
* prompt textbox;
* prompt initially populated from the class name;
* reset prompt to class name;
* Detect all objects.

Detect all objects provides prompt-based, class-aware detection within the current image or video frame.

Detect all objects is available on images and individual video frames. It uses shortcut **S**, opens a **Create annotations** popup with a SAM 3 badge, and requires an active class with bounding-box, polygon, mask, or cuboid geometry. The prompt defaults to the class name and accepts up to 300 characters. Each call is scoped to the current image or current frame, can return up to a configured maximum number of candidates, and consumes one AI-inference quota unit when successful.

Annotators review proposed objects before accepting them. Prompt changes can alter the candidate distribution even when the ontology class name remains the same, so teams can standardize prompts in their labeling instructions.

### Find Similar

Find Similar is a contextual header action, not a drawing tool. It becomes available when the selected object is a box, polygon, or mask. It searches the current image or current video frame, exposes a confidence threshold, removes strong overlaps with existing annotations, and holds new predictions for review before commitment. It does not search across the full dataset and does not consume AI-inference quota.

Find Similar can return multiple candidates at the selected confidence threshold, with **Clear** and **Accept all** actions available. Seed quality, visual repetition, clutter, scale, and threshold affect the result, so operators review each candidate set in context before acceptance.

### Auto-Tracking and interpolation

Auto-Tracking predicts later frames with machine learning; interpolation fills geometry between keyframes. Both results remain visible on the timeline for human review and correction.

### Magic Crop

Magic Crop runs model-assisted labeling inside a user-selected crop region. It is useful when the relevant objects occupy one part of a larger surface and whole-image detection would create unnecessary proposals. Each call uses AI-inference quota and still requires review before acceptance.

### AI captioning

AI-assisted captioning can propose text for the current item. The proposal should be reviewed against the item-property or captioning ontology rather than treated as an automatically accepted description.

### Batch auto-annotation

Unitlab supports batch auto-annotation for model-assisted labeling. Batch jobs apply model assistance across a selected data scope, while human review remains part of the annotation and workflow process.

### Model stages

A Model stage can be placed between human stages in a workflow. Model output can move forward to annotation or review, while rejected outcomes return through the configured correction route. Inference behavior follows the connected model’s input, output, and class mapping configuration.

### Introduce an assist safely

{% stepper %}
{% step %}

#### 1. Match the assist to the task

Choose Magic Touch, Detect all objects, Find Similar, tracking, interpolation, Magic Crop, captioning, batch auto-annotation, or a Model stage.
{% endstep %}

{% step %}

#### 2. Run a controlled sample

Start with one item or a small cohort before any bulk action.
{% endstep %}

{% step %}

#### 3. Inspect the proposal

Review class, geometry, relations, identity, temporal range, properties, and Item Properties.
{% endstep %}

{% step %}

#### 4. Correct in the native editor

Treat generated content as editable draft work.
{% endstep %}

{% step %}

#### 5. Measure error patterns

Record misses and false positives by source, class, model or assist version, and operation.
{% endstep %}

{% step %}

#### 6. Expand through workflow

Increase volume only after the corrected output is stable and a review route is active.
{% endstep %}
{% endstepper %}

### Decisions that affect production

| Decision             | Production guidance                                                                                              |
| -------------------- | ---------------------------------------------------------------------------------------------------------------- |
| Interactive vs batch | Use interactive assists for expert-in-the-loop correction; use batch only after the same behavior is calibrated. |
| Confidence           | Use as a triage or routing signal, not a substitute for the quality contract.                                    |
| Class mapping        | Validate every model-to-ontology mapping before production.                                                      |
| Failure              | Route failed or empty output visibly; never let it look like completed annotation.                               |

{% hint style="warning" %}
A faster proposal that increases reviewer correction or hides failure is not a production improvement.
{% endhint %}

### Continue the operating flow

* Keep assisted work inside the same ontology and workflow.
* Review systematic error cohorts in Project Data.
* Version material model or mapping changes.