An Azure Custom Vision alternative before the 2028 retirement
Azure AI Custom Vision retires on September 25, 2028, and calls to it fail after that. Pictograph trains classification, detection, and segmentation models on your images and serves them as GPU endpoints, or exports them for the edge.
Pictograph vs Azure Custom Vision
What a Custom Vision project does today, and where it goes on Pictograph.
| Feature | Pictograph | Custom Vision |
|---|---|---|
| Available after September 25, 2028 | No: retired, calls fail | |
| Image classification | ||
| Object detection | RF-DETR, YOLOX | |
| Instance segmentation and keypoints | – | |
| AI-assisted labeling | SAM3 point, box, and text prompts | Smart Labeler |
| Hosted prediction API | TensorRT on T4, L4, A100 | |
| Live video streams (RTSP, WebRTC) | – | |
| Run models on your own hardware | ONNX, TensorRT, ExecuTorch | Compact model exports |
Where Pictograph fits
Custom Vision projects that need a new home for the same job: label images, train a classifier or detector, and call it from an app or run it on a device. Pictograph adds SAM3 labeling, segmentation and keypoint models, TensorRT endpoints, and live RTSP and WebRTC video.
Where Azure leads
Staying inside Azure: your identity, networking, and billing do not change. Microsoft’s own successors are Azure Machine Learning AutoML, with a no-code studio close to Custom Vision’s, and generative options in Microsoft Foundry. If your models must run in your Azure tenant, start there.
Bring a Custom Vision project over
Convert its images and regions to a COCO file, then upload and import in two calls.
# Your Custom Vision images and regions, converted to a COCO file
pictograph images upload-directory parts-inspection ./images
pictograph annotations import-coco parts-inspection ./instances.json
# Missing labels? Draft them from a text prompt and review them in the app
pictograph sam3 predict part.jpg --text "scratch" --jsonfrom pictograph import Client
client = Client()
# Your Custom Vision images and regions, converted to a COCO file.
client.images.upload_from_directory(
dataset_name="parts-inspection",
directory="./images",
)
report = client.annotations.import_coco(
dataset_name="parts-inspection",
coco="./instances.json",
)
print(report.images_matched, "matched;", report.unmatched_files, "unmatched")The annotations reference also imports YOLO and Pascal VOC labels, and the images reference covers tagging images for classification.
Move off Custom Vision in four steps
Export while the API still answers, then train and serve on Pictograph.
-
Export your labels
Custom Vision’s training API lists every tagged image with its original-image URL and its regions as fractions of the image. Download the images and write the regions to a COCO file in pixels.
-
Import them
Upload the images to a Pictograph dataset and import the COCO file. Draft any missing labels with SAM3 and review them.
-
Train a model
Export the dataset and train a detector (RF-DETR or YOLOX) or a classifier on a managed GPU, then check its precision, recall, and mAP.
-
Deploy or export it
Serve it as a GPU endpoint for images or live video, or download ONNX, TensorRT, or ExecuTorch weights to replace a compact-model export at the edge.
The training guide walks through a run, deployments cover serving it, and local inference covers ONNX, TensorRT, and ExecuTorch on your own hardware.
Azure Custom Vision alternative FAQ
When is Azure Custom Vision being retired?
On September 25, 2028. Microsoft says calls to the service fail after that date and that existing customers are fully supported until then. It asked customers to plan their transition by September 25, 2026, so if you have not picked a replacement yet, you are behind Microsoft’s own schedule.
What does Microsoft recommend instead?
Microsoft lists three options: Azure Machine Learning AutoML, which trains classification and detection models with code or a no-code studio; generative models in Microsoft Foundry; and Azure Content Understanding, in preview, for generative image classification. If your models must stay inside Azure, start with AutoML. Pictograph is the option for teams that want labeling, training, and real-time serving in one product outside Azure.
How do I get my data out of Custom Vision?
Through the Custom Vision training API, before the retirement date. It returns each tagged image with its original-image URL and its regions, each a tag plus left, top, width, and height as fractions of the image. Multiply those by the image size to get pixel boxes, write them to a COCO file, then upload the images to Pictograph and import the file. Classification tags can be set on each image through the SDK or the app.
Can I still run my model on a device, like a Custom Vision compact model?
Yes. A model trained on Pictograph downloads as ONNX, and can be built as a TensorRT engine for NVIDIA hardware or an ExecuTorch file for phones and ARM boards. The Python SDK runs each of them with the same typed result, so the code that calls the model does not change between them.
What does Pictograph add over Custom Vision?
Instance segmentation and keypoint models, SAM3 labeling from a text prompt, TensorRT endpoints on T4, L4, and A100 GPUs, and live video: a deployment can pull an RTSP camera or take a WebRTC stream and return every frame’s results, with tracking, line counts, and dwell time in workflows. A CLI and a Claude Code plugin let a coding agent run the whole loop.
Other comparisons
Move your Custom Vision project
$5 of free compute every month. No card required.