SAM 3
Segment objects in images.

SAM 3 is Meta’s model for promptable segmentation in images and video. On AI Fabric, start with an image worker that turns a text description into object masks, boxes and scores your application can use.

€5 in free credits for every new customer.

Plan your deployment

AI Fabric access is opening gradually. GPU configuration is confirmed during onboarding.

Object masks for editing and annotation.

Turn an object description into a selection your application can use. Review the mask before using it to cut out a product, label training data or guide an edit.

Prepare product cutouts.

Ask for the product in a photograph, inspect the resulting mask and use it to build a transparent cutout. Keep edge correction available for reflective surfaces and fine details.

Assist dataset annotation.

Pre-fill candidate masks for an annotator to approve or refine. Store the image, prompt and model revision with each suggestion so your training data remains traceable.

Target an image edit.

Use a selected mask to tell a separate editor which region may change. Show the selection before launching the edit so the user can correct an overly broad result.

Model capabilities and upstream setup: SAM 3 publisher reference. Reviewed 2026-10-12.

Deployment preparation

Run text-prompted SAM 3 segmentation in Python.

The Python example performs text-prompted segmentation on a local image. Obtain access to the gated checkpoint and use the matching Transformers processor before preparing the worker.

Arrange authorized model access and keep the download token in the runtime’s secret configuration.
Load a representative image and inspect masks, bounding boxes and confidence scores from a simple prompt.
Serialize the approved output as your own mask asset and metadata contract, then test resizing and coordinate mapping.

AI Fabric preparation guide

python

import torch
from PIL import Image
from transformers import Sam3Model, Sam3Processor

model_id = "facebook/sam3"
model = Sam3Model.from_pretrained(model_id).to("cuda")
processor = Sam3Processor.from_pretrained(model_id)
image = Image.open("product.jpg").convert("RGB")
inputs = processor(
    images=image, text="bottle", return_tensors="pt",
).to("cuda")
with torch.inference_mode():
    predictions = model(**inputs)
result = processor.post_process_instance_segmentation(
    predictions, threshold=0.5, mask_threshold=0.5,
    target_sizes=inputs["original_sizes"].tolist(),
)[0]
print("Candidate objects:", len(result["masks"]))

Requires authorized checkpoint access, a compatible Transformers build, PyTorch and Pillow, plus product.jpg. The thresholds are example application choices; calibrate them on reviewed images.

GPU memory & performanceValidate peak memory using your actual image size and batch. A successful still-image run does not establish capacity for the separate video tracking workflow.

In your applicationPreserve original image dimensions. A mask rescaled with the wrong coordinate convention can select the wrong region even when the model prediction is useful.

AI Fabric availability & costRTX A4500 (20 GB) and RTX 2000 Ada (16 GB) are the GPU options. Access, runtime settings and billing are confirmed during onboarding. No Blade latency or throughput benchmark is published for this configuration.

Use SAM 3 masks to guide image editing.

Keep object selection and generation as separate jobs. Let the user inspect the mask, then pass the approved region to an image-editing model. Save both outputs so an unexpected edit can be traced to either selection or generation.

01 / InputImage + object description

02 / ProcessSAM 3 proposes masks

03 / ResultReview selection → edit

SAM 3
deployment FAQ.

Does SAM 3 generate a replacement background?

It produces segmentation outputs. A separate compositing or image-generation step is needed to replace a background or create new content.

Can a text prompt select several matching objects?

SAM 3 supports promptable concept segmentation. Inspect all returned masks and scores, then let the application decide whether to keep one object or multiple instances.

Can I reuse this worker for long videos unchanged?

No. Video processing has its own state, frame and memory requirements. Size and test that workflow separately before exposing it through the same service.

How much does running SAM 3 on AI Fabric cost?

Measure cost per processed image at your target resolution and batch size. Budget video separately: frame count, tracking state and output storage make it a different workload.

The GPU billing unit and idle policy are confirmed during onboarding. Managed Models API rates do not apply to this deployment.

Checkpoint: facebook/sam3. Access requirements and licence: SAM 3 publisher reference.

Browse all AI Fabric model guides ↗