---

name: "2d-workflows" description: "Use for StarDist 2D workflows: Config2D data contracts, training handoff, pretrained or local StarDist2D inference, normalization and axes, dense or sparse prediction, scaling, tiling, multiclass prediction, shape completion, threshold tuning, and block-wise large-image prediction." metadata: disco-role: operating disable-model-invocation: true license: BSD 3-Clause

StarDist 2D workflows

Use this sub-skill to prepare, train, load, diagnose, or run a 2D StarDist2D model. Keep the image/label contract, model configuration, normalization policy, and prediction parameters together in the handoff.

This route targets StarDist 0.9.2 at commit e80c6de700693bc228ed3c9ba1dc19c3785667ee. Its required baseline is CPU TensorFlow 2.x plus compiled CPU StarDist extensions. Network/model downloads, OpenCL/gputools, CUDA, and external integrations are optional and must be explicit.

Route by intent

• Use workflows for ordered preparation, bounded training pseudocode, pretrained/local inference, threshold tuning, and handoff contracts.
• Use api-reference for exact 0.9.2 signatures, defaults, model storage, input/output shapes, and export boundaries.
• Use multiclass when n_classes is not None or each detected object needs a class.
• Use large-data for n_tiles, scale, chunked output, or predict_instances_big.
• Use troubleshooting before changing axes, channels, normalization, thresholds, grids, blocks, or backends to recover from a failure.

Use evaluation-geometry for standalone geometry/NMS/matching. Use deployment-integration for console, BioImage.IO, ROI, Fiji, and QuPath work. Do not use this sub-skill for 3D model or ray selection. A ZYX/ZYXC volume is not a 2D image; select and document a plane as YX/YXC, or use 3d-workflows.

Operating route

1. Establish shape, dtype, axes, channels, model source, normalization, resources, and acceptance checks. In 2D, images are YX or channel-last YXC; labels are integer YX, background 0, and positive instance ids.
2. Inspect model.config.axes, n_channel_in, n_rays, grid, and n_classes. Never treat a successfully constructed untrained model as a substitute for failed pretrained/local weight loading.
3. Normalize exactly once. normalizer=None means already normalized; either supply normalized float data or a compatible csbdeep.data.Normalizer and record its statistics/channel policy.
4. Use bounded pseudocode only for training. Training changes checkpoints and logs, can consume substantial CPU/RAM, and is not bundled as a script here. Shape completion is a training choice, not an inference switch.
5. Start instance inference with default sparse=True. Use dense mode only for full maps/diagnostics; return_predict=True forces dense mode. Record probability/NMS thresholds, scale, and one tile count per input axis.
6. Escalate memory from spatial n_tiles to predict_instances_big. Big prediction additionally requires measured object-size bounds, sufficient context, grid-compatible blocks, and min_overlap + 2*context < block_size.
7. Verify label shape/dtype, polygon detail counts/finite values, and a known crop or fixture before batch use. Handoff model identity, config, axes, normalization, exact prediction options, output contract, optional-resource requirements, warnings, and unresolved suitability assumptions.

Evidence boundary

Native evidence was distilled from relative repository paths: stardist/models/model2d.py, stardist/models/base.py, stardist/models/__init__.py, stardist/big.py, stardist/data/__init__.py, README.md, the examples/2D/ and relevant examples/other2D/ notebooks, and tests/test_model2D.py, tests/test_stardist2D.py, tests/test_big.py, tests/test_utils.py, plus their 2D fixtures/loaders. These paths are evidence, not runtime dependencies; the bundled operating files do not import repository examples or tests.