napari-sam3-assistant 3.0.0

A napari plugin for interactive Segment Anything Model 3 (SAM3) segmentation using text, points, boxes, exemplar prompts, large-image ROI inference, mask operations, and 3D/video-like propagation

napari-sam3-assistant

napari-sam3-assistant is a napari plugin for interactive Segment Anything Model 3 (SAM3) segmentation using text, points, boxes, exemplar prompts, large-image ROI inference, point-based refinement, mask operations, and 3D/video-like propagation.

The plugin focuses on task-based segmentation workflows:

• 2D segmentation with text, box, point, and mask-style prompts
• 3D stack / video-like propagation from prompts on a selected slice or frame
• exemplar segmentation from Shapes ROI boxes
• text-based concept segmentation
• large OME-Zarr and TIFF segmentation through local ROI inference
• refinement with positive and negative point prompts
• downstream mask cleanup, merge, and export operations

SAM 3 is not bundled with this plugin. Install the SAM 3 backend and download the SAM 3 model files separately from Meta's Hugging Face repository.

Status

This project is under active development. The current widget supports local SAM 3 model loading, napari prompt collection, large-image ROI execution, downstream mask operations, background execution, and writing results back to napari layers.

Changelog

3.0.0

Major update compared with 2.0.0, focused on large-image segmentation, Step 7 mask operations, workflow modularization, and bug fixes.

Added:

• Optional large-image local inference mode for OME-Zarr, large TIFF, and similar very large images, including images on the order of 60000 x 50000 pixels when the data source can provide lazy ROI reads.
• Local ROI inference with selectable ROI sizes from 512 x 512 through 8192 x 8192.
• Active ROI overlay layer showing the current SAM3 local inference window in global image coordinates.
• Step 7 Mask Operations workflow for accepting, cleaning, merging, and exporting segmentation masks.
• Accepted-object saving with object name, class name, class value, append, and replace modes.
• Class-level merge workflow for accepted object layers.
• Mask cleanup tools for connected-component analysis, deleting selected components, removing small objects, filling holes, smoothing masks, keeping the largest object, and relabeling values.
• Final merge/export tools for semantic, instance, and binary output masks.
• Overlap handling during final mask merge with priority, selection-order, component-size, and background rules.
• Mask export to TIFF, PNG, and NumPy .npy.
• Task-runner modules for image, refinement, and video workflows to keep the main widget easier to maintain.

Changed:

• Large-image ROI choices now include 4096 x 4096 and 8192 x 8192.
• README now documents the large-image workflow and Step 7 mask operations.
• Result writing handles ROI-local outputs and maps them back into global image coordinates.
• Prompt collection and coordinate utilities support ROI-local conversion for large-image workflows.

Fixed:

• Improved coordinate handling for local ROI segmentation on large images.
• Improved preview/result layer handling for ROI-local outputs.
• Added tests for ROI extraction, prompt localization, global result mapping, and Step 7 UI presence.

2.0.0

Major update focused on SAM3.1 support, clearer model selection, live refinement, and result handling.

Added:

• SAM3.1 video multiplex support through sam3.1_multiplex.pt.
• Explicit Model type selector:
  • SAM3.0 2D/3D/video
  • SAM3.1 video multiplex
• Model-type-aware validation so SAM3.0 and SAM3.1 folders are not confused.
• SAM3.1 routing for 3D/video propagation through the multiplex video predictor.
• Automatic task guidance for SAM3.1: Load 2D Model is disabled and the task is set to 3D/video propagation.
• Two-column step-based widget layout:
  • left column: Model Setup, Task, Layers
  • right column: Prompt Tools, Run, Results, and collapsible Status
• Muted professional widget theme for lower eye strain during long napari sessions.
• Compact activity indicator in Run showing model execution, propagation, refinement, idle, and failure states.
• Results table showing Layer, Prompt, Frame, Object ID, Score, and Area.
• Detection threshold control for SAM3 grounding, useful when text prompts return no candidates.
• More visible text prompt input with Enter bound to Run Preview.
• Text prompt cleanup that sends short model-facing phrases such as myelin ring instead of instruction text such as segment all the myelin ring.
• Automatic lower-threshold retry for text segmentation when the first pass returns zero objects.
• Batch mode for running the same 2D prompt setup across all open image layers, with separate preview and saved label layers per image.
• Multi-text batch mode: enter one text concept per line and run each prompt independently against the selected image or all image layers.
• Results actions:
  • Clear Results
  • Copy Clipboard
  • Export CSV
• Label-value merge controls for converting multiple label IDs into one class value.
• Stable object-ID label mapping for propagated video/stack results.
• Remembered model type, model directory, and device selection through Qt settings.
• Safer CUDA error reporting for unsupported GPU kernel architectures.

Live refinement improvements:

• Refinement (live point correction) mode now arms live point refinement immediately after Create Prompt Layer.
• The first point starts live refinement; the first run may take longer if the model is lazy-loaded.
• SAM3 preview labels is pre-created for live refinement so napari does not switch users away from the points layer after the first result.
• After each live refinement update, the active layer returns to SAM3 points in add mode.
• Next point mode affects only future points.
• T toggles next point mode only and does not rerun refinement.
• Shift+T flips selected point polarity, or the latest point if none is selected, and reruns refinement.
• Existing dot colors now change only when stored point properties["polarity"] changes.
• Apply mode to selected points still edits selected existing points and reruns preview.

Changed:

• Backend / Model Setup was renamed to Model Setup.
• Napari Layers was renamed to Layers.
• Batch all image layers can be used independently from multi-text prompts.
• Load Image Model was renamed to Load 2D Model.
• Lazy-load on run was renamed to Load model when running.
• Text prompts no longer require a prompt layer.
• Text prompt submission can be run with Enter without moving to the Run Preview button.
• Preview clearing removes generated preview layers only and keeps prompts, saved labels, and loaded models.
• pytorch_model.bin was removed from documented and validated model-file names.

Current model support:

• SAM3.0 weights support 2D image tasks and 3D/video propagation.
• SAM3.1 sam3.1_multiplex.pt supports 3D/video propagation.
• SAM3.1 is not currently routed through the plugin's 2D image model loader.

1.0.0

Initial SAM3 Assistant plugin foundation.

Added:

• Local SAM3 backend adapter for napari workflows.
• Task-based UI for:
  • 2D segmentation
  • 3D stack/video-like propagation
  • exemplar segmentation
  • text segmentation
  • refinement with positive and negative prompts
• Prompt collection from napari Points, Shapes, Labels, and text input.
• Box prompts from Shapes layers.
• Labels-layer mask prompts.
• Text prompts for concept segmentation.
• Preview outputs as napari Labels, Image, and Shapes layers.
• Saved label outputs through Save Result as Labels.
• Background worker execution to keep the napari UI responsive.
• Channel-axis handling for grayscale, RGB/RGBA, channel-first, and stack-like data.
• Basic model-directory validation for local SAM3 files.

Requirements

• Python >=3.12
• napari >=0.5
• SAM 3 Python package importable as sam3
• PyTorch and torchvision installed for your platform
• A local SAM 3 checkpoint directory containing:
  • config.json
  • processor_config.json
  • one weight file such as sam3.pt, model.safetensors, or sam3.1_multiplex.pt

If CUDA is not available or not compatible, select CPU in the widget.

Setup

Important for Windows users

If you installed the standalone napari desktop app from the website, Plugin Manager may not work for this plugin because that app uses a different Python environment.

For Windows, use a new Conda environment with Python 3.12 and install napari, SAM 3, and napari-sam3-assistant there.

Windows install in PowerShell

# 1) install Miniforge first from:
# https://conda-forge.org/download/

# 2) open Miniforge Prompt or PowerShell after Miniforge is installed


conda create -n napari-sam3 python=3.12 -y
conda activate napari-sam3

python -m pip install --upgrade pip wheel
python -m pip install "setuptools<82"
python -m pip install "napari[all]"

# choose one:
python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
# or
python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

python -m pip install --no-cache-dir sam3
python -m pip install einops triton-windows pycocotools
python -m pip install napari-sam3-assistant

napari

Download SAM 3 model files

Download the model files from:

• https://huggingface.co/facebook/sam3
• https://huggingface.co/facebook/sam3.1

These repositories are gated, so you must request or accept access first.

After approval, download the files manually from the Files and versions tab.

A reference screenshot of the file list is shown below:

Model folder

The model files can be stored in any folder you want.

Examples:

D:\models\sam3
D:\models\sam3_1

The plugin only needs the correct folder path.

Example SAM3.0 folder:

D:\models\sam3\
  config.json
  processor_config.json
  sam3.pt

model.safetensors is also supported as a SAM3 3.0 weight file.

Example SAM3.1 folder:

D:\models\sam3_1\
  config.json
  processor_config.json
  sam3.1_multiplex.pt

Current model support:

• SAM3.0 weights: 2D image tasks and 3D/video propagation
• SAM3.1 sam3.1_multiplex.pt: 3D/video propagation through the SAM3.1 multiplex video predictor
• SAM3.1 is not currently routed through the plugin's 2D image model loader

Device rule

• If you installed CPU-only PyTorch, choose CPU
• If you installed CUDA PyTorch and torch.cuda.is_available() is True, you may choose CUDA

For developers

If you want editable installs and git pull, use:

git clone https://github.com/facebookresearch/sam3.git
cd sam3
python -m pip install --no-cache-dir -e .

git clone https://github.com/wulinteousa2-hash/napari-sam3-assistant.git
cd napari-sam3-assistant
python -m pip install -e .

Basic Workflow

1. Open an image in napari.
2. Open Plugins > SAM3 Assistant.
3. Select the image in Layers > Image.
4. Select a task.
5. Create a prompt layer if the task needs one.
6. Click Run Preview.
7. Inspect SAM3 preview labels, SAM3 preview masks, or SAM3 preview boxes.
8. Click Save Result as Labels to keep the result.

Use Clear Preview to remove generated preview layers without deleting prompts or saved labels.

Large-Image Local Inference

Large-image mode is optional and off by default. When it is off, the plugin keeps the existing full-image inference path.

Use this mode for OME-Zarr, large TIFF, and similar large images where sending the full selected plane to SAM3 is too expensive.

Workflow:

1. Set up the normal task and prompt type.
2. Enable Enable large-image local inference in Step 2. Task.
3. Choose a local ROI size:

512 x 512
1024 x 1024
2048 x 2048
4096 x 4096
8192 x 8192

4. Add a point or box prompt.
5. Click Run Preview, or add points in live refinement mode.

ROI behavior:

• Point prompts use the latest point as the ROI anchor.
• Box prompts use the box center and keep the box inside the local inference window when possible.
• Refinement uses the latest point as the ROI anchor.
• Text-only prompts keep the full-image path in this first pass unless a point or box anchor is also available.
• If a new point or box stays inside the current ROI, the same ROI is reused.
• If a new point or box falls outside the current ROI, the ROI is rebuilt around the new prompt.

The active ROI is shown as:

SAM3 active ROI

SAM3 receives only the local ROI image data. Returned labels and boxes are written back into global image coordinates in the normal preview layers.

Status messages report:

Large-image mode OFF: full-image inference.
Large-image mode ON: local ROI inference (WIDTH x HEIGHT).
Active ROI bounds: y=Y0:Y1, x=X0:X1.

Batch Multiple 2D Images

Use Batch all image layers when several open 2D images should receive the same prompt setup.

Workflow:

1. Open multiple images in napari.
2. Configure a 2D task such as text, box, exemplar, or labels-mask segmentation.
3. Add the prompt once.
4. Enable Batch all image layers in Step 3. Layers.
5. Click Run Preview.

Each source image gets its own output layers:

SAM3 preview labels [image name]
SAM3 preview masks [image name]
SAM3 preview boxes [image name]

Save Result as Labels saves each batch preview labels layer separately:

SAM3 saved labels [image name]

Batch mode is intended for 2D image tasks. It is disabled for live refinement and 3D/video propagation because those workflows depend on one active image/session.

Multi-Text Batch Mode

Use Batch text prompts when you want each text concept to run independently instead of writing one combined phrase such as cat and dog.

Workflow:

1. Set Task to Text segmentation.
2. Enter one concept per line in Batch text prompts:

cat
dog
person

3. Leave Batch all image layers off to run all prompts on the selected image only.
4. Enable Batch all image layers to run every prompt on every open image.
5. Click Run Preview.

Outputs include both image and prompt:

SAM3 preview labels [Image 1 - cat]
SAM3 preview labels [Image 1 - dog]
SAM3 preview labels [Image 2 - cat]
SAM3 preview labels [Image 2 - dog]

The Results table includes a Prompt column. Object IDs are scoped to each image-prompt result, so Object ID 1 for Image 1 - cat is separate from Object ID 1 for Image 1 - dog.

Tasks

Text Segmentation

Use text to segment all matching instances of a concept.

Workflow:

1. Set Task to Text segmentation.
2. Leave Prompt type as Text only.
3. Enter a short phrase, for example:

cell
nucleus
myelin
myelin sheath

4. Keep Detection threshold near the default 0.35, or lower it if the result is empty.
5. Press Enter in the text prompt field or click Run Preview.

No prompt layer is needed for text segmentation. Create Prompt Layer is not required.

Text prompts usually work better as short noun phrases than instructions. Prefer myelin sheath over segment all the myelin rings. The plugin strips common instruction prefixes before sending the prompt to SAM3, but microscopy-specific language can still be difficult for the model.

If the result says objects=0, SAM3 ran but did not return masks above threshold. Try a shorter noun phrase, lower Detection threshold, or use a box/exemplar prompt for structures that are visually clear but not well recognized by text.

2D Segmentation With Boxes

Use boxes to identify the target object or concept.

Workflow:

1. Set Task to 2D segmentation.
2. Set Prompt type to Box.
3. Click Create Prompt Layer.
4. Draw one or more rectangles in the SAM3 boxes Shapes layer.
5. Click Run Preview.

The output appears in preview layers.

Exemplar Segmentation

Use example ROIs to segment similar objects.

Workflow:

1. Set Task to Exemplar segmentation.
2. Set Prompt type to Box.
3. Click Create Prompt Layer.
4. Draw boxes around one or more example objects.
5. Click Run Preview.

The local SAM 3 image API exposes visual exemplars through geometric box prompts. The plugin stores ROI metadata, but inference currently passes exemplar ROIs as SAM 3 visual box prompts.

Refinement With Positive and Negative Points

Use points to correct a result.

Workflow:

1. Set Task to Refinement (live point correction).
2. Set Prompt type to Points (positive/negative).
3. Click Create Prompt Layer.
4. Choose Positive and add points on regions to include.
5. Choose Negative and add points on regions to exclude.
6. Click Run Preview.

This is useful after a text, box, or exemplar preview is close but not correct.

Labels Mask Prompt

Use a napari Labels layer as a mask-style prompt.

Workflow:

1. Set a task that supports mask prompts.
2. Set Prompt type to Labels mask.
3. Click Create Prompt Layer.
4. Paint non-zero pixels in SAM3 mask prompt.
5. Click Run Preview.

3D Stack / Video Propagation

Treat a stack as video-like data and propagate a prompt through frames or slices.

Workflow:

1. Open a stack in napari.
2. Set Task to 3D/video propagation.
3. Select the target frame or slice in napari.
4. Create a prompt layer and add prompts on that frame.
5. Choose propagation direction:
   • both
   • forward
   • backward
6. Click Run Preview or Propagate Stack/Video.

Preview output is written to:

SAM3 propagated preview labels

Saved output is written to:

SAM3 saved propagated labels

The current SAM 3 video predictor backend is CUDA-only. CPU mode is supported for 2D/image workflows, not 3D/video propagation.

Channel Axis

Channel axis tells the plugin which data axis is color/channel.

Default:

-1

Use -1 for grayscale images and normal RGB/RGBA images. The plugin auto-detects trailing RGB/RGBA axes of size 3 or 4.

Examples:

(H, W)          -> -1
(H, W, 3)      -> -1
(H, W, 4)      -> -1
(Z, H, W)      -> -1
(C, H, W)      -> 0
(Z, C, H, W)   -> 1
(T, C, H, W)   -> 1
(Z, H, W, C)   -> 3

Leave it at -1 unless your image has an explicit multi-channel microscopy dimension.

Output Layers

Preview layers:

SAM3 preview labels
SAM3 preview masks
SAM3 preview boxes
SAM3 propagated preview labels

Saved layers:

SAM3 saved labels
SAM3 saved propagated labels

Buttons:

• Validate: check the selected SAM 3 model directory.
• Load 2D Model: load the 2D/image model.
• Load 3D/Video Model: load the video propagation model.
• Run Preview: run the selected task.
• Clear Preview: remove generated preview layers only.
• Save Result as Labels: copy preview labels into saved labels.
• Cancel: stop a running worker.
• Unload: unload the SAM3 model from memory.

Results table:

Layer | Prompt | Frame | Object ID | Score | Area

• Layer: source image layer.
• Prompt: text prompt used for text and multi-text results. For non-text workflows this is -.
• Frame: propagated frame or slice index. For 2D results this is -.
• Object ID: SAM3 object ID when available, otherwise a generated label ID.
• Score: SAM3 confidence/probability when returned by the backend.
• Area: number of mask pixels for that object in the displayed 2D plane or frame.

Results actions:

• Clear Results: clear the table only.
• Copy Clipboard: copy tab-separated results, including headers, for pasting into Excel or statistics software.
• Export CSV: save the results table to a CSV file.

Label-value merge:

Use this when multiple SAM3 objects should become the same class value in a Labels layer. For example, if labels 3, 4, 5, and 6 are all the same biological class, set:

Values to replace: 3,4,5,6
New value: 3

Then click Merge Label Values. The selected Labels layer is updated in place.

Step 7 Mask Operations

Step 7 provides downstream mask operations for turning SAM3 previews into curated masks for analysis or training data.

Tabs:

• Accepted Objects: save a preview Labels layer as a named accepted object with class metadata, append it to an existing accepted layer, or replace an existing accepted layer.
• Class Merge: merge selected accepted-object layers into a class working mask.
• Mask Cleanup: analyze connected components, delete selected components, remove small objects, fill holes, smooth masks, keep the largest component, and relabel values.
• Final Merge / Export: merge cleaned class masks into semantic, instance, or binary final masks, choose overlap handling, and export TIFF, PNG, or NumPy .npy files.

The mask operations panel works on napari Labels layers, including SAM3 preview and saved label layers.

ARM64, CUDA, and DGX Spark

For ARM64 systems such as NVIDIA DGX Spark / GB10:

• Use Python 3.12 or newer.
• Keep the NVIDIA driver and CUDA stack current.
• Install a PyTorch/torchvision build that supports your GPU architecture.
• Use CPU mode for reliable 2D execution if CUDA kernels are unavailable.
• Use explicit CUDA only when testing a compatible GPU build.

Check PyTorch GPU support:

python - <<'PY'
import torch
print("torch:", torch.__version__)
print("torch cuda runtime:", torch.version.cuda)
print("cuda available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("device:", torch.cuda.get_device_name(0))
    print("capability:", torch.cuda.get_device_capability(0))
    print("arch list:", torch.cuda.get_arch_list())
PY

GB10 reports compute capability 12.1 (sm_121). If your PyTorch build does not include compatible kernels, you may see:

CUDA error: no kernel image is available for execution on the device
nvrtc: error: invalid value for --gpu-architecture

The plugin does not compile PyTorch, torchvision, or SAM 3 CUDA extensions.

Troubleshooting

No mask appears and status says objects=0

SAM 3 returned no detections above threshold. Try:

• a shorter text prompt
• a more common concept phrase
• a lower Detection threshold
• box or exemplar prompts
• CPU mode if the CUDA path is unstable

CUDA kernel image error

Error:

CUDA error: no kernel image is available for execution on the device

The GPU is visible, but at least one required CUDA kernel was not built for the device architecture. Use CPU, or install compatible PyTorch/torchvision/SAM 3 builds.

Invalid GPU architecture

Error:

nvrtc: error: invalid value for --gpu-architecture

The installed PyTorch CUDA runtime cannot compile for the detected GPU. Use CPU or install a build that supports the GPU.

BFloat16 conversion errors

The plugin converts SAM3 bfloat16 outputs to float32 before writing NumPy-backed napari layers. If you still see dtype errors, restart napari after changing device mode and run again.

Text prompt creates no layer

That is expected. Text segmentation does not need a prompt layer. Enter text and click Run Preview.

Development

Install in editable mode:

pip install -e .

Run tests:

PYTHONPATH=src pytest -q

The test suite covers coordinate mapping, prompt collection, adapter utility behavior, and static widget UI checks. It does not download SAM 3 weights.

References

• SAM 3 repository: https://github.com/facebookresearch/sam3
• SAM 3 model files: https://huggingface.co/facebook/sam3
• PyTorch installation selector: https://pytorch.org/get-started/locally/

Acknowledgement

The demo image was provided by the Electron Microscopy Core Facility at Houston Methodist Research Institute

License

MIT. See the project license file.