SAM v2.1 Hiera Large

SAM v2.1 (Segment Anything Model 2.1) represents the next evolution in promptable visual segmentation by Meta AI, delivering more accurate and efficient mask generation for a wide range of image types and contexts.

Model Information

• Architecture: SAM v2.1 extends the original SAM framework with optimized Hiera-based encoders for higher accuracy and speed.
• Flexible Outputs: Supports overlay images, polygon coordinates, COCO RLE encodings, and individual PNG masks.

How to Use the Model

1. Upload an image (JPG, PNG, or WebP) — either as a URL or base64 string.
2. Configure output options:
   • make_overlay – Generate a visual overlay of masks on the input image.
   • save_polygons – Return polygon coordinates for each segmented region.
   • save_rle – Export COCO RLE encodings for mask data.
   • save_pngs – Save individual masks as PNGs in a ZIP file.
3. Adjust advanced parameters for quality and precision:
   • points_per_side, points_per_batch, pred_iou_thresh, stability_score_thresh, min_mask_region_area, nms_iou_thresh, max_masks, polygon_epsilon, tile_size, tile_stride.
4. Click “Generate” to obtain segmentation results.

Use Cases

• Assisted Image Labeling: Speeds up dataset annotation with automatic mask proposals.
• AR/VR Applications: Enables precise object isolation for immersive environments.
• Autonomous Vehicles: Supports accurate perception and obstacle detection.
• Environmental Monitoring: Segments satellite and aerial imagery for analysis.
• Industrial & Sonar Imaging: Identifies regions of interest in complex visual data.