How facebookresearch/segment-anything works
SAM is a promptable image-segmentation model: users load a checkpoint, provide an image and point/box/mask prompts, and receive masks plus quality scores; automatic generation samples prompts and filters masks across an image. The graph also preserves the documented command-line, ONNX export, and browser-demo capabilities. Core predictor/model relationships are supported by excerpts and the source index; the demo implementation and some utility wiring were not sampled, so those links are documented or left unverified.
Main entry points: sam.py SAM orchestration · predictor.py Prompt predictor · mask_decoder.py Mask decoder · automatic_mask_generator.py Automatic generator
Components 17
User APIs
- Model registry
segment_anything/build_sam.py - Prompt predictor
segment_anything/predictor.py - Image transforms
segment_anything/utils/transforms.py
SAM Model
- SAM orchestration
segment_anything/modeling/sam.py - ViT image encoder
segment_anything/modeling/image_encoder.py - Prompt encoder
segment_anything/modeling/prompt_encoder.py - Mask decoder
segment_anything/modeling/mask_decoder.py - Two-way transformer
segment_anything/modeling/transformer.py - Model layers
segment_anything/modeling/common.py
Automatic Masks
- Automatic generator
segment_anything/automatic_mask_generator.py - Mask CLI
scripts/amg.py - Mask utilities
segment_anything/utils/amg.py
Export And Demo
- ONNX export CLI
scripts/export_onnx_model.py - ONNX wrapper
segment_anything/utils/onnx.py - Browser demo
demo/src - ONNX Runtime
Other components
- Image user
Connections 23
- Image user → Model registry: loads checkpoint
README.md - Model registry → SAM orchestration: builds model
segment_anything/build_sam.py - Image user → Prompt predictor: sets image and prompts
README.md - Prompt predictor → Image transforms: transforms inputs
segment_anything/predictor.py - Prompt predictor → SAM orchestration: requests masks
segment_anything/predictor.py - SAM orchestration → ViT image encoder: encodes image
segment_anything/modeling/sam.py - SAM orchestration → Prompt encoder: encodes prompts
segment_anything/modeling/sam.py - SAM orchestration → Mask decoder: decodes masks
segment_anything/modeling/sam.py - Mask decoder → Two-way transformer: attends
- ViT image encoder → Model layers: uses layers
segment_anything/modeling/image_encoder.py - Mask decoder → Model layers: uses layers
segment_anything/modeling/mask_decoder.py - Two-way transformer → Model layers: uses layers
segment_anything/modeling/transformer.py - Image user → Automatic generator: requests all masks
README.md - Automatic generator → Prompt predictor: predicts sampled prompts
segment_anything/automatic_mask_generator.py - Automatic generator → Mask utilities: filters and encodes
segment_anything/automatic_mask_generator.py - Image user → Mask CLI: runs image command
README.md - Mask CLI → Automatic generator: generates masks
- Image user → ONNX export CLI: requests export
README.md - ONNX export CLI → ONNX wrapper: exports wrapper
scripts/export_onnx_model.py - ONNX wrapper → Prompt encoder: uses prompt encoding
segment_anything/utils/onnx.py - ONNX wrapper → Mask decoder: decodes masks
segment_anything/utils/onnx.py - Browser demo → ONNX Runtime: runs exported model
README.md - Image user → Browser demo: uses web demo
README.md