Auto-annotate with SAM3
A click, a box, or a phrase becomes pixel-precise masks, on one image or thousands. Every result is an editable draft you review before you export and train.
From a prompt to reviewed labels
No setup and no model to wire up. SAM3 drafts the masks; you decide which are right.
-
Prompt it
Click a point, drag a box, type a phrase, or label one example.
-
SAM3 drafts
A managed GPU returns pixel-precise polygons or boxes for every match.
-
You review
Filter to the least certain drafts and fix them; approve each image or send it back.
-
Export and train
Export the dataset to COCO, YOLO or ten more formats, and train a model on the export.
One model, every way to prompt it
Every mode runs the same SAM3 model and returns the same editable polygons and boxes, from the app, Python, or the CLI.
from pictograph import Client
client = Client()
# Positive points include, negative points exclude
result = client.auto_annotate.point(
dataset_name="Market Shelf",
image_filename="img-001.jpg",
x=412,
y=280,
name="tomato",
negative_points=[(90, 120)],
)from pictograph import Client
client = Client()
# Draw one box around an object
result = client.auto_annotate.box(
dataset_name="Market Shelf",
image_filename="img-001.jpg",
box={"x": 360, "y": 220, "w": 120, "h": 140},
name="tomato",
)from pictograph import Client
client = Client()
# Label by concept, no training required
result = client.auto_annotate.text(
dataset_name="Market Shelf",
image_filename="img-001.jpg",
text_prompt="bell pepper",
)# SAM3 text prompts over many images, saved to each one
pictograph auto-annotate batch "Market Shelf" \
--images img-001.jpg,img-002.jpg,img-003.jpg \
--classes "tomato:polygon,bell pepper:polygon"
# Or pre-label with a model you trained
pictograph auto-annotate batch "Market Shelf" \
--images img-004.jpg,img-005.jpg --model produce-detectorBeyond a single prompt
Label a whole dataset from one example, from your own model, or at a scale where small objects matter.
Manual annotation is always free. Draw boxes, rotated boxes, polygons, polylines, and keypoints by hand, and fix any draft the same way.
Keep exploring
Auto-annotation FAQ
What is SAM3 auto-annotation?
SAM3 auto-annotation turns a point, a box, or a text prompt into polygon or bounding-box labels in one call. Click an object, drag a box, or type a class name, and the model returns precise masks you can save as annotations.
Can I auto-label a whole dataset at once?
Yes. Batch auto-annotation runs SAM3 text prompts across many images and classes in one job, and Annotate similar uses one labeled object as an example to find the matching objects across a folder or the whole dataset. Overlapping detections of the same class are merged, and a batch job bills compute credits for the GPU time it uses. It all runs on every plan, Free included.
Should I train on the auto-labels directly?
Review them first: auto-labels are drafts. Filter to the drafts the model was least sure of and fix them; the review stage lets you approve an image or send it back with a note. Then export the dataset and train on the export.
Does auto-annotation support text prompts?
Yes. Type a phrase like "car" or "traffic light" and SAM3 finds every matching region in the image, returning a mask for each. It works on one image interactively or across a dataset in batch.
Can I pre-label with my own model?
Yes. Once you have trained a model on Pictograph, a batch job can run it over new images instead of SAM3, adding any of its classes the dataset does not have yet. You review and correct the results like any other annotation.
How accurate is SAM3 segmentation?
SAM3 returns pixel-precise masks, including holes inside an object. Refine any result by adding positive or negative points, or by editing the polygon vertices directly.