[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6ADVHF9CKKF5H7SSMBNPZ7Y":3,"skill-same:01M6ADVHF9CKKF5H7SSMBNPZ7Y":118},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":26,"publishedAt":49,"linkOnly":50,"markdownContent":51,"markdownBody":52,"frontMatter":53,"refName":66,"commitSha":67,"skillDirectory":68,"forksCount":69,"repositoryHtmlUrl":70,"repositoryHomepageUrl":71,"ownerHtmlUrl":72,"ownerBlogUrl":73,"ownerDescription":74,"updatedAt":75,"install":76,"resources":89},"01M6ADVHF9CKKF5H7SSMBNPZ7Y","multimodal-segment-anything",null,"Foundation model for image segmentation with zero-shot transfer. Use when you need to segment any object in images using points, boxes, or masks as prompts, or automatically generate all object masks in an image.","davila7\u002Fclaude-code-templates","davila7","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F6216945?v=4","01M6ADEWWFPS191NFM0Y1EX1JH","Daniel Avila","01M6ADJHGM78PHQRD3XZTFDVKK",33481,"MIT","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-segment-anything\u002FSKILL.md",[18,23],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000016","图像与多媒体","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",[27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48],"multimodal","multimodal segment anything","segment anything","segment","vision","when","for","orchestra","you","Zero-Shot","boxes","as","images","image","computer","model","segmentation","Transformers","ZeRO","transfer","generate","research","2026-11-10 10:12:30",false,"---\nname: multimodal-segment-anything\ndescription: Foundation model for image segmentation with zero-shot transfer. Use when you need to segment any object in images using points, boxes, or masks as prompts, or automatically generate all object masks in an image.\nlicense: MIT\nmetadata:\n  version: 1.0.0\n  author: Orchestra Research\n  tags: [Multimodal, Image Segmentation, Computer Vision, SAM, Zero-Shot]\n  dependencies: [segment-anything, transformers>=4.30.0, torch>=1.7.0]\n---\n\n# Segment Anything Model (SAM)\n\nComprehensive guide to using Meta AI's Segment Anything Model for zero-shot image segmentation.\n\n## When to use SAM\n\n**Use SAM when:**\n- Need to segment any object in images without task-specific training\n- Building interactive annotation tools with point\u002Fbox prompts\n- Generating training data for other vision models\n- Need zero-shot transfer to new image domains\n- Building object detection\u002Fsegmentation pipelines\n- Processing medical, satellite, or domain-specific images\n\n**Key features:**\n- **Zero-shot segmentation**: Works on any image domain without fine-tuning\n- **Flexible prompts**: Points, bounding boxes, or previous masks\n- **Automatic segmentation**: Generate all object masks automatically\n- **High quality**: Trained on 1.1 billion masks from 11 million images\n- **Multiple model sizes**: ViT-B (fastest), ViT-L, ViT-H (most accurate)\n- **ONNX export**: Deploy in browsers and edge devices\n\n**Use alternatives instead:**\n- **YOLO\u002FDetectron2**: For real-time object detection with classes\n- **Mask2Former**: For semantic\u002Fpanoptic segmentation with categories\n- **GroundingDINO + SAM**: For text-prompted segmentation\n- **SAM 2**: For video segmentation tasks\n\n## Quick start\n\n### Installation\n\n```bash\n# From GitHub\npip install git+https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything.git\n\n# Optional dependencies\npip install opencv-python pycocotools matplotlib\n\n# Or use HuggingFace transformers\npip install transformers\n```\n\n### Download checkpoints\n\n```bash\n# ViT-H (largest, most accurate) - 2.4GB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_h_4b8939.pth\n\n# ViT-L (medium) - 1.2GB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_l_0b3195.pth\n\n# ViT-B (smallest, fastest) - 375MB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_b_01ec64.pth\n```\n\n### Basic usage with SamPredictor\n\n```python\nimport numpy as np\nfrom segment_anything import sam_model_registry, SamPredictor\n\n# Load model\nsam = sam_model_registry[\"vit_h\"](checkpoint=\"sam_vit_h_4b8939.pth\")\nsam.to(device=\"cuda\")\n\n# Create predictor\npredictor = SamPredictor(sam)\n\n# Set image (computes embeddings once)\nimage = cv2.imread(\"image.jpg\")\nimage = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\npredictor.set_image(image)\n\n# Predict with point prompts\ninput_point = np.array([[500, 375]])  # (x, y) coordinates\ninput_label = np.array([1])  # 1 = foreground, 0 = background\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_point,\n    point_labels=input_label,\n    multimask_output=True  # Returns 3 mask options\n)\n\n# Select best mask\nbest_mask = masks[np.argmax(scores)]\n```\n\n### HuggingFace Transformers\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import SamModel, SamProcessor\n\n# Load model and processor\nmodel = SamModel.from_pretrained(\"facebook\u002Fsam-vit-huge\")\nprocessor = SamProcessor.from_pretrained(\"facebook\u002Fsam-vit-huge\")\nmodel.to(\"cuda\")\n\n# Process image with point prompt\nimage = Image.open(\"image.jpg\")\ninput_points = [[[450, 600]]]  # Batch of points\n\ninputs = processor(image, input_points=input_points, return_tensors=\"pt\")\ninputs = {k: v.to(\"cuda\") for k, v in inputs.items()}\n\n# Generate masks\nwith torch.no_grad():\n    outputs = model(**inputs)\n\n# Post-process masks to original size\nmasks = processor.image_processor.post_process_masks(\n    outputs.pred_masks.cpu(),\n    inputs[\"original_sizes\"].cpu(),\n    inputs[\"reshaped_input_sizes\"].cpu()\n)\n```\n\n## Core concepts\n\n### Model architecture\n\n```\nSAM Architecture:\n┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐\n│  Image Encoder  │────▶│ Prompt Encoder  │────▶│  Mask Decoder   │\n│     (ViT)       │     │ (Points\u002FBoxes)  │     │ (Transformer)   │\n└─────────────────┘     └─────────────────┘     └─────────────────┘\n        │                       │                       │\n   Image Embeddings      Prompt Embeddings         Masks + IoU\n   (computed once)       (per prompt)             predictions\n```\n\n### Model variants\n\n| Model | Checkpoint | Size | Speed | Accuracy |\n|-------|------------|------|-------|----------|\n| ViT-H | `vit_h` | 2.4 GB | Slowest | Best |\n| ViT-L | `vit_l` | 1.2 GB | Medium | Good |\n| ViT-B | `vit_b` | 375 MB | Fastest | Good |\n\n### Prompt types\n\n| Prompt | Description | Use Case |\n|--------|-------------|----------|\n| Point (foreground) | Click on object | Single object selection |\n| Point (background) | Click outside object | Exclude regions |\n| Bounding box | Rectangle around object | Larger objects |\n| Previous mask | Low-res mask input | Iterative refinement |\n\n## Interactive segmentation\n\n### Point prompts\n\n```python\n# Single foreground point\ninput_point = np.array([[500, 375]])\ninput_label = np.array([1])\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_point,\n    point_labels=input_label,\n    multimask_output=True\n)\n\n# Multiple points (foreground + background)\ninput_points = np.array([[500, 375], [600, 400], [450, 300]])\ninput_labels = np.array([1, 1, 0])  # 2 foreground, 1 background\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_points,\n    point_labels=input_labels,\n    multimask_output=False  # Single mask when prompts are clear\n)\n```\n\n### Box prompts\n\n```python\n# Bounding box [x1, y1, x2, y2]\ninput_box = np.array([425, 600, 700, 875])\n\nmasks, scores, logits = predictor.predict(\n    box=input_box,\n    multimask_output=False\n)\n```\n\n### Combined prompts\n\n```python\n# Box + points for precise control\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375]]),\n    point_labels=np.array([1]),\n    box=np.array([400, 300, 700, 600]),\n    multimask_output=False\n)\n```\n\n### Iterative refinement\n\n```python\n# Initial prediction\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375]]),\n    point_labels=np.array([1]),\n    multimask_output=True\n)\n\n# Refine with additional point using previous mask\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375], [550, 400]]),\n    point_labels=np.array([1, 0]),  # Add background point\n    mask_input=logits[np.argmax(scores)][None, :, :],  # Use best mask\n    multimask_output=False\n)\n```\n\n## Automatic mask generation\n\n### Basic automatic segmentation\n\n```python\nfrom segment_anything import SamAutomaticMaskGenerator\n\n# Create generator\nmask_generator = SamAutomaticMaskGenerator(sam)\n\n# Generate all masks\nmasks = mask_generator.generate(image)\n\n# Each mask contains:\n# - segmentation: binary mask\n# - bbox: [x, y, w, h]\n# - area: pixel count\n# - predicted_iou: quality score\n# - stability_score: robustness score\n# - point_coords: generating point\n```\n\n### Customized generation\n\n```python\nmask_generator = SamAutomaticMaskGenerator(\n    model=sam,\n    points_per_side=32,          # Grid density (more = more masks)\n    pred_iou_thresh=0.88,        # Quality threshold\n    stability_score_thresh=0.95,  # Stability threshold\n    crop_n_layers=1,             # Multi-scale crops\n    crop_n_points_downscale_factor=2,\n    min_mask_region_area=100,    # Remove tiny masks\n)\n\nmasks = mask_generator.generate(image)\n```\n\n### Filtering masks\n\n```python\n# Sort by area (largest first)\nmasks = sorted(masks, key=lambda x: x['area'], reverse=True)\n\n# Filter by predicted IoU\nhigh_quality = [m for m in masks if m['predicted_iou'] > 0.9]\n\n# Filter by stability score\nstable_masks = [m for m in masks if m['stability_score'] > 0.95]\n```\n\n## Batched inference\n\n### Multiple images\n\n```python\n# Process multiple images efficiently\nimages = [cv2.imread(f\"image_{i}.jpg\") for i in range(10)]\n\nall_masks = []\nfor image in images:\n    predictor.set_image(image)\n    masks, _, _ = predictor.predict(\n        point_coords=np.array([[500, 375]]),\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n    all_masks.append(masks)\n```\n\n### Multiple prompts per image\n\n```python\n# Process multiple prompts efficiently (one image encoding)\npredictor.set_image(image)\n\n# Batch of point prompts\npoints = [\n    np.array([[100, 100]]),\n    np.array([[200, 200]]),\n    np.array([[300, 300]])\n]\n\nall_masks = []\nfor point in points:\n    masks, scores, _ = predictor.predict(\n        point_coords=point,\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n    all_masks.append(masks[np.argmax(scores)])\n```\n\n## ONNX deployment\n\n### Export model\n\n```bash\npython scripts\u002Fexport_onnx_model.py \\\n    --checkpoint sam_vit_h_4b8939.pth \\\n    --model-type vit_h \\\n    --output sam_onnx.onnx \\\n    --return-single-mask\n```\n\n### Use ONNX model\n\n```python\nimport onnxruntime\n\n# Load ONNX model\nort_session = onnxruntime.InferenceSession(\"sam_onnx.onnx\")\n\n# Run inference (image embeddings computed separately)\nmasks = ort_session.run(\n    None,\n    {\n        \"image_embeddings\": image_embeddings,\n        \"point_coords\": point_coords,\n        \"point_labels\": point_labels,\n        \"mask_input\": np.zeros((1, 1, 256, 256), dtype=np.float32),\n        \"has_mask_input\": np.array([0], dtype=np.float32),\n        \"orig_im_size\": np.array([h, w], dtype=np.float32)\n    }\n)\n```\n\n## Common workflows\n\n### Workflow 1: Annotation tool\n\n```python\nimport cv2\n\n# Load model\npredictor = SamPredictor(sam)\npredictor.set_image(image)\n\ndef on_click(event, x, y, flags, param):\n    if event == cv2.EVENT_LBUTTONDOWN:\n        # Foreground point\n        masks, scores, _ = predictor.predict(\n            point_coords=np.array([[x, y]]),\n            point_labels=np.array([1]),\n            multimask_output=True\n        )\n        # Display best mask\n        display_mask(masks[np.argmax(scores)])\n```\n\n### Workflow 2: Object extraction\n\n```python\ndef extract_object(image, point):\n    \"\"\"Extract object at point with transparent background.\"\"\"\n    predictor.set_image(image)\n\n    masks, scores, _ = predictor.predict(\n        point_coords=np.array([point]),\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n\n    best_mask = masks[np.argmax(scores)]\n\n    # Create RGBA output\n    rgba = np.zeros((image.shape[0], image.shape[1], 4), dtype=np.uint8)\n    rgba[:, :, :3] = image\n    rgba[:, :, 3] = best_mask * 255\n\n    return rgba\n```\n\n### Workflow 3: Medical image segmentation\n\n```python\n# Process medical images (grayscale to RGB)\nmedical_image = cv2.imread(\"scan.png\", cv2.IMREAD_GRAYSCALE)\nrgb_image = cv2.cvtColor(medical_image, cv2.COLOR_GRAY2RGB)\n\npredictor.set_image(rgb_image)\n\n# Segment region of interest\nmasks, scores, _ = predictor.predict(\n    box=np.array([x1, y1, x2, y2]),  # ROI bounding box\n    multimask_output=True\n)\n```\n\n## Output format\n\n### Mask data structure\n\n```python\n# SamAutomaticMaskGenerator output\n{\n    \"segmentation\": np.ndarray,  # H×W binary mask\n    \"bbox\": [x, y, w, h],        # Bounding box\n    \"area\": int,                 # Pixel count\n    \"predicted_iou\": float,      # 0-1 quality score\n    \"stability_score\": float,    # 0-1 robustness score\n    \"crop_box\": [x, y, w, h],    # Generation crop region\n    \"point_coords\": [[x, y]],    # Input point\n}\n```\n\n### COCO RLE format\n\n```python\nfrom pycocotools import mask as mask_utils\n\n# Encode mask to RLE\nrle = mask_utils.encode(np.asfortranarray(mask.astype(np.uint8)))\nrle[\"counts\"] = rle[\"counts\"].decode(\"utf-8\")\n\n# Decode RLE to mask\ndecoded_mask = mask_utils.decode(rle)\n```\n\n## Performance optimization\n\n### GPU memory\n\n```python\n# Use smaller model for limited VRAM\nsam = sam_model_registry[\"vit_b\"](checkpoint=\"sam_vit_b_01ec64.pth\")\n\n# Process images in batches\n# Clear CUDA cache between large batches\ntorch.cuda.empty_cache()\n```\n\n### Speed optimization\n\n```python\n# Use half precision\nsam = sam.half()\n\n# Reduce points for automatic generation\nmask_generator = SamAutomaticMaskGenerator(\n    model=sam,\n    points_per_side=16,  # Default is 32\n)\n\n# Use ONNX for deployment\n# Export with --return-single-mask for faster inference\n```\n\n## Common issues\n\n| Issue | Solution |\n|-------|----------|\n| Out of memory | Use ViT-B model, reduce image size |\n| Slow inference | Use ViT-B, reduce points_per_side |\n| Poor mask quality | Try different prompts, use box + points |\n| Edge artifacts | Use stability_score filtering |\n| Small objects missed | Increase points_per_side |\n\n## References\n\n- **[Advanced Usage](references\u002Fadvanced-usage.md)** - Batching, fine-tuning, integration\n- **[Troubleshooting](references\u002Ftroubleshooting.md)** - Common issues and solutions\n\n## Resources\n\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.02643\n- **Demo**: https:\u002F\u002Fsegment-anything.com\n- **SAM 2 (Video)**: https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything-2\n- **HuggingFace**: https:\u002F\u002Fhuggingface.co\u002Ffacebook\u002Fsam-vit-huge\n","# Segment Anything Model (SAM)\n\nComprehensive guide to using Meta AI's Segment Anything Model for zero-shot image segmentation.\n\n## When to use SAM\n\n**Use SAM when:**\n- Need to segment any object in images without task-specific training\n- Building interactive annotation tools with point\u002Fbox prompts\n- Generating training data for other vision models\n- Need zero-shot transfer to new image domains\n- Building object detection\u002Fsegmentation pipelines\n- Processing medical, satellite, or domain-specific images\n\n**Key features:**\n- **Zero-shot segmentation**: Works on any image domain without fine-tuning\n- **Flexible prompts**: Points, bounding boxes, or previous masks\n- **Automatic segmentation**: Generate all object masks automatically\n- **High quality**: Trained on 1.1 billion masks from 11 million images\n- **Multiple model sizes**: ViT-B (fastest), ViT-L, ViT-H (most accurate)\n- **ONNX export**: Deploy in browsers and edge devices\n\n**Use alternatives instead:**\n- **YOLO\u002FDetectron2**: For real-time object detection with classes\n- **Mask2Former**: For semantic\u002Fpanoptic segmentation with categories\n- **GroundingDINO + SAM**: For text-prompted segmentation\n- **SAM 2**: For video segmentation tasks\n\n## Quick start\n\n### Installation\n\n```bash\n# From GitHub\npip install git+https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything.git\n\n# Optional dependencies\npip install opencv-python pycocotools matplotlib\n\n# Or use HuggingFace transformers\npip install transformers\n```\n\n### Download checkpoints\n\n```bash\n# ViT-H (largest, most accurate) - 2.4GB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_h_4b8939.pth\n\n# ViT-L (medium) - 1.2GB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_l_0b3195.pth\n\n# ViT-B (smallest, fastest) - 375MB\nwget https:\u002F\u002Fdl.fbaipublicfiles.com\u002Fsegment_anything\u002Fsam_vit_b_01ec64.pth\n```\n\n### Basic usage with SamPredictor\n\n```python\nimport numpy as np\nfrom segment_anything import sam_model_registry, SamPredictor\n\n# Load model\nsam = sam_model_registry[\"vit_h\"](checkpoint=\"sam_vit_h_4b8939.pth\")\nsam.to(device=\"cuda\")\n\n# Create predictor\npredictor = SamPredictor(sam)\n\n# Set image (computes embeddings once)\nimage = cv2.imread(\"image.jpg\")\nimage = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\npredictor.set_image(image)\n\n# Predict with point prompts\ninput_point = np.array([[500, 375]])  # (x, y) coordinates\ninput_label = np.array([1])  # 1 = foreground, 0 = background\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_point,\n    point_labels=input_label,\n    multimask_output=True  # Returns 3 mask options\n)\n\n# Select best mask\nbest_mask = masks[np.argmax(scores)]\n```\n\n### HuggingFace Transformers\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import SamModel, SamProcessor\n\n# Load model and processor\nmodel = SamModel.from_pretrained(\"facebook\u002Fsam-vit-huge\")\nprocessor = SamProcessor.from_pretrained(\"facebook\u002Fsam-vit-huge\")\nmodel.to(\"cuda\")\n\n# Process image with point prompt\nimage = Image.open(\"image.jpg\")\ninput_points = [[[450, 600]]]  # Batch of points\n\ninputs = processor(image, input_points=input_points, return_tensors=\"pt\")\ninputs = {k: v.to(\"cuda\") for k, v in inputs.items()}\n\n# Generate masks\nwith torch.no_grad():\n    outputs = model(**inputs)\n\n# Post-process masks to original size\nmasks = processor.image_processor.post_process_masks(\n    outputs.pred_masks.cpu(),\n    inputs[\"original_sizes\"].cpu(),\n    inputs[\"reshaped_input_sizes\"].cpu()\n)\n```\n\n## Core concepts\n\n### Model architecture\n\n```\nSAM Architecture:\n┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐\n│  Image Encoder  │────▶│ Prompt Encoder  │────▶│  Mask Decoder   │\n│     (ViT)       │     │ (Points\u002FBoxes)  │     │ (Transformer)   │\n└─────────────────┘     └─────────────────┘     └─────────────────┘\n        │                       │                       │\n   Image Embeddings      Prompt Embeddings         Masks + IoU\n   (computed once)       (per prompt)             predictions\n```\n\n### Model variants\n\n| Model | Checkpoint | Size | Speed | Accuracy |\n|-------|------------|------|-------|----------|\n| ViT-H | `vit_h` | 2.4 GB | Slowest | Best |\n| ViT-L | `vit_l` | 1.2 GB | Medium | Good |\n| ViT-B | `vit_b` | 375 MB | Fastest | Good |\n\n### Prompt types\n\n| Prompt | Description | Use Case |\n|--------|-------------|----------|\n| Point (foreground) | Click on object | Single object selection |\n| Point (background) | Click outside object | Exclude regions |\n| Bounding box | Rectangle around object | Larger objects |\n| Previous mask | Low-res mask input | Iterative refinement |\n\n## Interactive segmentation\n\n### Point prompts\n\n```python\n# Single foreground point\ninput_point = np.array([[500, 375]])\ninput_label = np.array([1])\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_point,\n    point_labels=input_label,\n    multimask_output=True\n)\n\n# Multiple points (foreground + background)\ninput_points = np.array([[500, 375], [600, 400], [450, 300]])\ninput_labels = np.array([1, 1, 0])  # 2 foreground, 1 background\n\nmasks, scores, logits = predictor.predict(\n    point_coords=input_points,\n    point_labels=input_labels,\n    multimask_output=False  # Single mask when prompts are clear\n)\n```\n\n### Box prompts\n\n```python\n# Bounding box [x1, y1, x2, y2]\ninput_box = np.array([425, 600, 700, 875])\n\nmasks, scores, logits = predictor.predict(\n    box=input_box,\n    multimask_output=False\n)\n```\n\n### Combined prompts\n\n```python\n# Box + points for precise control\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375]]),\n    point_labels=np.array([1]),\n    box=np.array([400, 300, 700, 600]),\n    multimask_output=False\n)\n```\n\n### Iterative refinement\n\n```python\n# Initial prediction\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375]]),\n    point_labels=np.array([1]),\n    multimask_output=True\n)\n\n# Refine with additional point using previous mask\nmasks, scores, logits = predictor.predict(\n    point_coords=np.array([[500, 375], [550, 400]]),\n    point_labels=np.array([1, 0]),  # Add background point\n    mask_input=logits[np.argmax(scores)][None, :, :],  # Use best mask\n    multimask_output=False\n)\n```\n\n## Automatic mask generation\n\n### Basic automatic segmentation\n\n```python\nfrom segment_anything import SamAutomaticMaskGenerator\n\n# Create generator\nmask_generator = SamAutomaticMaskGenerator(sam)\n\n# Generate all masks\nmasks = mask_generator.generate(image)\n\n# Each mask contains:\n# - segmentation: binary mask\n# - bbox: [x, y, w, h]\n# - area: pixel count\n# - predicted_iou: quality score\n# - stability_score: robustness score\n# - point_coords: generating point\n```\n\n### Customized generation\n\n```python\nmask_generator = SamAutomaticMaskGenerator(\n    model=sam,\n    points_per_side=32,          # Grid density (more = more masks)\n    pred_iou_thresh=0.88,        # Quality threshold\n    stability_score_thresh=0.95,  # Stability threshold\n    crop_n_layers=1,             # Multi-scale crops\n    crop_n_points_downscale_factor=2,\n    min_mask_region_area=100,    # Remove tiny masks\n)\n\nmasks = mask_generator.generate(image)\n```\n\n### Filtering masks\n\n```python\n# Sort by area (largest first)\nmasks = sorted(masks, key=lambda x: x['area'], reverse=True)\n\n# Filter by predicted IoU\nhigh_quality = [m for m in masks if m['predicted_iou'] > 0.9]\n\n# Filter by stability score\nstable_masks = [m for m in masks if m['stability_score'] > 0.95]\n```\n\n## Batched inference\n\n### Multiple images\n\n```python\n# Process multiple images efficiently\nimages = [cv2.imread(f\"image_{i}.jpg\") for i in range(10)]\n\nall_masks = []\nfor image in images:\n    predictor.set_image(image)\n    masks, _, _ = predictor.predict(\n        point_coords=np.array([[500, 375]]),\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n    all_masks.append(masks)\n```\n\n### Multiple prompts per image\n\n```python\n# Process multiple prompts efficiently (one image encoding)\npredictor.set_image(image)\n\n# Batch of point prompts\npoints = [\n    np.array([[100, 100]]),\n    np.array([[200, 200]]),\n    np.array([[300, 300]])\n]\n\nall_masks = []\nfor point in points:\n    masks, scores, _ = predictor.predict(\n        point_coords=point,\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n    all_masks.append(masks[np.argmax(scores)])\n```\n\n## ONNX deployment\n\n### Export model\n\n```bash\npython scripts\u002Fexport_onnx_model.py \\\n    --checkpoint sam_vit_h_4b8939.pth \\\n    --model-type vit_h \\\n    --output sam_onnx.onnx \\\n    --return-single-mask\n```\n\n### Use ONNX model\n\n```python\nimport onnxruntime\n\n# Load ONNX model\nort_session = onnxruntime.InferenceSession(\"sam_onnx.onnx\")\n\n# Run inference (image embeddings computed separately)\nmasks = ort_session.run(\n    None,\n    {\n        \"image_embeddings\": image_embeddings,\n        \"point_coords\": point_coords,\n        \"point_labels\": point_labels,\n        \"mask_input\": np.zeros((1, 1, 256, 256), dtype=np.float32),\n        \"has_mask_input\": np.array([0], dtype=np.float32),\n        \"orig_im_size\": np.array([h, w], dtype=np.float32)\n    }\n)\n```\n\n## Common workflows\n\n### Workflow 1: Annotation tool\n\n```python\nimport cv2\n\n# Load model\npredictor = SamPredictor(sam)\npredictor.set_image(image)\n\ndef on_click(event, x, y, flags, param):\n    if event == cv2.EVENT_LBUTTONDOWN:\n        # Foreground point\n        masks, scores, _ = predictor.predict(\n            point_coords=np.array([[x, y]]),\n            point_labels=np.array([1]),\n            multimask_output=True\n        )\n        # Display best mask\n        display_mask(masks[np.argmax(scores)])\n```\n\n### Workflow 2: Object extraction\n\n```python\ndef extract_object(image, point):\n    \"\"\"Extract object at point with transparent background.\"\"\"\n    predictor.set_image(image)\n\n    masks, scores, _ = predictor.predict(\n        point_coords=np.array([point]),\n        point_labels=np.array([1]),\n        multimask_output=True\n    )\n\n    best_mask = masks[np.argmax(scores)]\n\n    # Create RGBA output\n    rgba = np.zeros((image.shape[0], image.shape[1], 4), dtype=np.uint8)\n    rgba[:, :, :3] = image\n    rgba[:, :, 3] = best_mask * 255\n\n    return rgba\n```\n\n### Workflow 3: Medical image segmentation\n\n```python\n# Process medical images (grayscale to RGB)\nmedical_image = cv2.imread(\"scan.png\", cv2.IMREAD_GRAYSCALE)\nrgb_image = cv2.cvtColor(medical_image, cv2.COLOR_GRAY2RGB)\n\npredictor.set_image(rgb_image)\n\n# Segment region of interest\nmasks, scores, _ = predictor.predict(\n    box=np.array([x1, y1, x2, y2]),  # ROI bounding box\n    multimask_output=True\n)\n```\n\n## Output format\n\n### Mask data structure\n\n```python\n# SamAutomaticMaskGenerator output\n{\n    \"segmentation\": np.ndarray,  # H×W binary mask\n    \"bbox\": [x, y, w, h],        # Bounding box\n    \"area\": int,                 # Pixel count\n    \"predicted_iou\": float,      # 0-1 quality score\n    \"stability_score\": float,    # 0-1 robustness score\n    \"crop_box\": [x, y, w, h],    # Generation crop region\n    \"point_coords\": [[x, y]],    # Input point\n}\n```\n\n### COCO RLE format\n\n```python\nfrom pycocotools import mask as mask_utils\n\n# Encode mask to RLE\nrle = mask_utils.encode(np.asfortranarray(mask.astype(np.uint8)))\nrle[\"counts\"] = rle[\"counts\"].decode(\"utf-8\")\n\n# Decode RLE to mask\ndecoded_mask = mask_utils.decode(rle)\n```\n\n## Performance optimization\n\n### GPU memory\n\n```python\n# Use smaller model for limited VRAM\nsam = sam_model_registry[\"vit_b\"](checkpoint=\"sam_vit_b_01ec64.pth\")\n\n# Process images in batches\n# Clear CUDA cache between large batches\ntorch.cuda.empty_cache()\n```\n\n### Speed optimization\n\n```python\n# Use half precision\nsam = sam.half()\n\n# Reduce points for automatic generation\nmask_generator = SamAutomaticMaskGenerator(\n    model=sam,\n    points_per_side=16,  # Default is 32\n)\n\n# Use ONNX for deployment\n# Export with --return-single-mask for faster inference\n```\n\n## Common issues\n\n| Issue | Solution |\n|-------|----------|\n| Out of memory | Use ViT-B model, reduce image size |\n| Slow inference | Use ViT-B, reduce points_per_side |\n| Poor mask quality | Try different prompts, use box + points |\n| Edge artifacts | Use stability_score filtering |\n| Small objects missed | Increase points_per_side |\n\n## References\n\n- **[Advanced Usage](references\u002Fadvanced-usage.md)** - Batching, fine-tuning, integration\n- **[Troubleshooting](references\u002Ftroubleshooting.md)** - Common issues and solutions\n\n## Resources\n\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.02643\n- **Demo**: https:\u002F\u002Fsegment-anything.com\n- **SAM 2 (Video)**: https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fsegment-anything-2\n- **HuggingFace**: https:\u002F\u002Fhuggingface.co\u002Ffacebook\u002Fsam-vit-huge\n",{"name":5,"license":15,"metadata":54,"description":7},{"tags":55,"author":60,"version":61,"dependencies":62},[56,57,58,59,36],"Multimodal","Image Segmentation","Computer Vision","SAM","Orchestra Research","1.0.0",[63,64,65],"segment-anything","transformers>=4.30.0","torch>=1.7.0","main","2b749c82b5d11fac39d9a6e021ce3f5b8a541bd1","cli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-segment-anything",3931,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates","https:\u002F\u002Faitmpl.com","https:\u002F\u002Fgithub.com\u002Fdavila7","https:\u002F\u002Fdanielavila.me","Building AI dev tools with LLMs","2026-11-14 06:14:46",{"source":8,"skillName":5,"commands":77,"prompt":87,"zipAvailable":88},[78,81,84],{"tool":79,"command":80},"npx","npx skills add davila7\u002Fclaude-code-templates --skill multimodal-segment-anything",{"tool":82,"command":83},"bunx","bunx skills add davila7\u002Fclaude-code-templates --skill multimodal-segment-anything",{"tool":85,"command":86},"pnpm","pnpm dlx skills add davila7\u002Fclaude-code-templates --skill multimodal-segment-anything","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-segment-anything\u002FSKILL.md",true,[90,96,100,104,108,114],{"path":91,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":94,"sourceHtmlUrl":95},"FORMS.md","REFERENCE","参考文档",14006,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdocument-processing\u002Fpdf-processing-pro\u002FFORMS.md",{"path":97,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":98,"sourceHtmlUrl":99},"OCR.md",2828,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdocument-processing\u002Fpdf-processing-pro\u002FOCR.md",{"path":101,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":102,"sourceHtmlUrl":103},"references\u002Fadvanced-usage.md",15803,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-segment-anything\u002Freferences\u002Fadvanced-usage.md",{"path":105,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":106,"sourceHtmlUrl":107},"references\u002Ftroubleshooting.md",11914,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-segment-anything\u002Freferences\u002Ftroubleshooting.md",{"path":109,"resourceType":110,"resourceTypeDesc":111,"sizeBytes":112,"sourceHtmlUrl":113},"scripts\u002Fanalyze_form.py","SCRIPT","脚本",8684,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdocument-processing\u002Fpdf-processing-pro\u002Fscripts\u002Fanalyze_form.py",{"path":115,"resourceType":92,"resourceTypeDesc":93,"sizeBytes":116,"sourceHtmlUrl":117},"TABLES.md",14911,"https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdocument-processing\u002Fpdf-processing-pro\u002FTABLES.md",[119,145,172,205,240,264],{"skillNo":120,"name":121,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":126,"publishedAt":49},"01M70NHBTQ8BV0A2SCAZXEPHCX","wealthville-pool-scoring","Score DeFi liquidity pools before providing liquidity. Returns a 0-100 pool score and an Enter\u002FHold\u002FExit\u002FReduce\u002FAvoid verdict with calibrated confidence for ~68,800 Solana pools and 575 EVM pools, plus a public miss-inclusive track record. Use when asked whether a pool is worth LPing into, which pools look best right now, or how accurate the scores have actually been.","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fweb-data\u002Fwealthville-pool-scoring\u002FSKILL.md",[125],{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},[127,128,129,130,131,33,132,133,134,32,135,136,137,138,139,140,141,142,143,144],"scoring","pool","wealthville pool scoring","liquidity","which","exit","now","use","reduce","with","record","track","verdict","evm","solana","how","public","look",{"skillNo":146,"name":147,"nameZh":6,"description":148,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":149,"categories":150,"keywords":152,"publishedAt":49},"01M70MB6VR8ERJ4EM8D5KHBMEV","technical-debt-auditor","Scans a codebase and produces a quantified technical debt report with a prioritized remediation plan. Analyzes dead code, TODO\u002FFIXME\u002FHACK markers, dependency staleness, complexity hotspots, code duplication signals, test coverage gaps, unused exports, and outdated patterns. Use this skill when the user asks to audit their codebase for tech debt, wants a health check of their code quality, says things like \"where should I focus refactoring\", \"find code smells\", \"audit my codebase\", \"what technical debt do we have\", \"prioritize cleanup work\", or asks about code quality metrics. Also trigger when the user mentions \"code rot\", \"legacy cleanup\", \"refactoring priorities\", or wants to understand the overall health of a project before a major effort. Trigger even if the user doesn't say \"technical debt\" explicitly — any request to systematically assess code quality and identify improvement priorities belongs here.","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdevelopment\u002Ftechnical-debt-auditor\u002FSKILL.md",[151],{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},[153,154,155,156,157,158,159,160,161,162,163,32,164,165,166,167,168,169,170,171],"technical debt","debt","technical","technical debt auditor","auditor","coverage","complexity","markers","project","cleanup","unused","code","todo","work","plan","focus","signals","audit","dependency",{"skillNo":173,"name":174,"nameZh":6,"description":175,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":176,"categories":177,"keywords":187,"publishedAt":49},"01M70KPYH9QNHCNME5GJQGE22K","readonly-sql-guard","Make an agent's database access provably read-only. Use when an agent runs caller-supplied SQL against a production\u002FERP\u002Fanalytics database (a \"query\" MCP tool, text-to-SQL, an admin console), or when reviewing DB-access code for write\u002FDDL\u002Fside-effect\u002Ftransaction-escape risks. Checks what a statement CALLS, not just its shape.","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fdatabase\u002Freadonly-sql-guard\u002FSKILL.md",[178,181,182],{"categoryNo":179,"categoryName":180,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000010","后端与接口",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":183,"categoryName":184,"taxonomyType":185,"taxonomyTypeDesc":186},"01JBSEEDCAT000000000000033","数据分析师","PROFESSION","适用职业",[188,189,190,191,192,193,194,33,195,196,197,198,199,200,201,134,202,203,32,204],"readonly","guard","sql guard","sql","readonly sql guard","erp","effect","read-only","db","just","tool","calls","only","MCP","access","database","query",{"skillNo":206,"name":207,"nameZh":6,"description":208,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":209,"licenseFileUrl":210,"sourceHtmlUrl":211,"categories":212,"keywords":216,"publishedAt":49},"01M6AFY6F7CC2BBZP1QD6VTNVC","web-artifacts-builder","Suite of tools for creating elaborate, multi-component claude.ai HTML artifacts using modern frontend web technologies (React, Tailwind CSS, shadcn\u002Fui). Use for complex artifacts requiring state management, routing, or shadcn\u002Fui components - not for simple single-file HTML\u002FJSX artifacts.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Futilities\u002Fweb-artifacts-builder\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Futilities\u002Fweb-artifacts-builder\u002FSKILL.md",[213],{"categoryNo":214,"categoryName":215,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000009","前端开发",[217,218,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234,235,236,237,238,33,239],"artifacts builder","web","web artifacts builder","artifacts","builder","state","routing","state management","component","css","frontend","ui","Tools","file","components","shadcn","react","Tailwind","suite","jsx","html","single-file","ui components",{"skillNo":241,"name":242,"nameZh":6,"description":243,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":209,"licenseFileUrl":244,"sourceHtmlUrl":245,"categories":246,"keywords":251,"publishedAt":49},"01M6AFWF33SFV9Z3SF4PEEE4BF","webapp-testing","Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fsecurity\u002Fwebapp-testing\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fsecurity\u002Fwebapp-testing\u002FSKILL.md",[247,250],{"categoryNo":248,"categoryName":249,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",{"categoryNo":214,"categoryName":215,"taxonomyType":21,"taxonomyTypeDesc":22},[252,253,254,255,256,218,257,227,228,258,259,260,261,33,262,136,263],"webapp testing","testing","webapp","browser","playwright","applications","debugging","toolkit","logs","screenshots","local","behavior",{"skillNo":265,"name":266,"nameZh":6,"description":267,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":209,"licenseFileUrl":268,"sourceHtmlUrl":269,"categories":270,"keywords":278,"publishedAt":49},"01M6AF67Q67JJJ5V51JR2P52KA","discernment-nudge","After you give a substantive answer or draft that the user may act on — advice or recommendations, drafted artifacts such as goals, plans, pitches, proposals, or emails, estimates or projections, analysis or interpretation of data, factual claims they may rely on, or a multi-step argument — invoke this skill BEFORE finalizing your reply and then, if it applies, append 2-3 short follow-up questions, each tied to something specific in what you just produced, that help the user check key facts, probe the reasoning or assumptions, and notice missing context. Do this at most once per conversation. Skip it when the user asked a trivial how-to or simple lookup, wants a purely educational explanation, asked you only to format, convert, or assemble a file from content they provided, is writing code they will run, is doing creative writing or casual chat, or already asked you to double-check, cite, or review — the skill file explains these boundaries and the exact output format.","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fproductivity\u002Fdiscernment-nudge\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fdavila7\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fproductivity\u002Fdiscernment-nudge\u002FSKILL.md",[271,274,275],{"categoryNo":272,"categoryName":273,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":276,"categoryName":277,"taxonomyType":185,"taxonomyTypeDesc":186},"01JBSEEDCAT000000000000036","内容创作者",[279,280,281,282,283,284,285,286,220,287,230,164,288,289,290,291,292,200,293,32,35,294,295,296,297,142,298],"discernment nudge","nudge","interpretation","help","run","reasoning","data","writing","lookup","content","review","conversation","short","up","invoke","analysis","claims","argument","assemble","recommendations"]