Model Sample Image Export

Do not use:

• incomplete runs
• partially written checkpoints
• outputs with empty/degenerate masks
• outputs driven by NaNs or obviously broken predictions

Required Workflow

1. Identify candidate checkpoints/runs in results/.
2. Verify the run is complete enough to trust.
3. If verification fails, schedule a few runs to train the model on a few categories.
4. Generate predictions from the checkpoint/run.
5. Inspect output quality before publishing images.
6. Save the selected images into docs/source/images/<model>/results/.
7. Update README/docs references.

Preferred Output Layout

• docs/source/images/<model>/results/0.png
• docs/source/images/<model>/results/1.png
• docs/source/images/<model>/results/2.png

If you have fewer than 3 trustworthy images, train the model on a few more categories to generate more sample images.

README Update Pattern

Preferred pattern:

### Sample Results

![Sample Result 1](/docs/source/images/<model>/results/0.png "Sample Result 1")

Repeat for additional images.

Docs Update Pattern

Preferred docs-page pattern:

    ## Sample Results

    ```{eval-rst}
    .. image:: ../../../../../images/<model>/results/0.png
    ```

Validation Rules

Before publishing an image:

1. Check that the referenced file exists.
2. Check that the image is visually plausible.
3. Check that the mask/anomaly region is not obviously wrong.
4. Check that the sample came from a trained or otherwise valid completed run.
5. If a model/category output is degenerate, exclude it and say so explicitly.

Reviewer checklist

• Check run completeness.
• Check image quality.
• Check exported file existence.
• Check README and docs references.

Repo-Specific Notes

• In this repo, some completed checkpoints can still produce bad masks.
• If generic visualization helpers fail, derive a narrow exporter for the specific model/run.
• Keep exporter scripts focused and traceable to the chosen checkpoints.
• When in doubt, prefer fewer trustworthy sample images over a full set of misleading ones.