Move from Roboflow to Pictograph in two steps.
Bring your datasets over with the built-in connector - images and annotations come across, and Roboflow’s COCO export converts to Pictograph JSON automatically. Start free, no credit card.
Import with the Roboflow connector
Two steps from the SDK, CLI, or app. The source key is used only to fetch your data.
from pictograph import Client
client = Client() # reads PICTOGRAPH_API_KEY
# 1. Validate your Roboflow key and list its datasets.
result = client.connectors.validate(
provider="roboflow",
api_key="rf_your_workspace_key",
)
for ds in result.datasets:
print(ds.id, ds.name, ds.image_count)
# 2. Import the ones you want. Images and annotations come across, with
# Roboflow's COCO export converted to Pictograph JSON automatically.
job = client.connectors.import_(
provider="roboflow",
api_key="rf_your_workspace_key",
datasets=result.datasets,
wait=True,
)
print(job.import_id, job.status)Prefer the app? Go to Datasets, then Import, then Roboflow, paste your workspace key, and pick the datasets to bring over. Either way the import runs as a tracked job, and imported images get embeddings, content tags, and thumbnails just like a fresh upload, so they are searchable immediately. Roboflow has no folder concept, so those datasets import flat; you can reorganize into directories afterward.
What comes across
Your images and their annotations - boxes, polygons, and keypoints - with class names preserved and any missing class colors filled in. From there you can re-annotate with SAM3, add classes, or export in any of seven formats.
What you redo
Trained model weights, as with any platform switch. That is usually a feature, not a chore: retrain on Pictograph in one call and choose YOLOX for speed or RF-DETR for accuracy on hard scenes - and compare the two on your own data.
Label, train, and deploy from code
Once your data is in, the rest of the loop is one API. Auto-annotate with SAM3 point, box, or text prompts - label "every forklift" across a directory in a single batch job - then train a detector or segmenter on a managed GPU and serve it as one authenticated /predict endpoint. The same REST API backs a typed Python SDK, a CLI, and a 37-tool agent registry, and your data stays private on every tier. See the connectors reference for the full import API and the YOLOX vs RF-DETR guide to pick a detector.
Migrating from Roboflow FAQ
Can I import my Roboflow datasets into Pictograph?
Yes. Pictograph has a built-in Roboflow connector. Validate your Roboflow workspace API key to list your datasets, then import the ones you want; the images and annotations are pulled directly, so there is no manual export-and-re-upload step. You can also import from the app: Datasets, then Import, then Roboflow.
Will my annotations come across?
Yes. Roboflow datasets export as COCO, and Pictograph converts that to its canonical annotation format automatically - bounding boxes, polygons, and keypoints, with your class names preserved. Any classes missing a color are assigned a distinct one on import.
What happens to my trained Roboflow models?
Trained model weights do not transfer between platforms - that is true of any switch, not specific to Pictograph. Once your data is imported you retrain on Pictograph, often in one call: pick YOLOX for speed or RF-DETR for accuracy and train from the imported export. Auto-annotation with SAM3 can also fill in labels you never finished in Roboflow.
Is my data private after I switch?
Yes. Every Pictograph tier, including the free plan, keeps your datasets and models private by default; you choose per entity when to make something public. On Roboflow’s free Public plan, uploaded data and trained models are visible to everyone on Roboflow Universe.
Is there a free way to try the migration?
Yes. Pictograph’s free tier is $0 with no credit card and includes monthly compute. Import a Roboflow dataset, re-annotate with SAM3 or train a model, and evaluate the result before committing. Manual annotation is always free.
Bring your Roboflow data over
$5/mo free compute. No payment method required.