How to Train Your Dice

Start with ordinary rolls

The Factory Baseline is ready on day one and may already recognize many of your dice. Personal training is optional; when your own model is ready, it can replace the Factory Baseline with recognition adapted to your exact dice, camera, tray, and lighting.

Before collecting images

• Confirm the correct active Dice Set.
• Keep the camera, tray, lighting, zoom, and Rolling Area stable.
• Recalibrate after meaningful setup changes.
• Start with one die type and keep every label accurate. Version 1.31.5 can collect examples from several dice of that selected type in one roll.

Step 1 — Natural Roll Training and multi-die training

1. Open Training → 1. Natural Roll Training.
2. Select one die type: d4, d6, d8, d10, d12, d00, or d20.
3. Roll one or more dice of that selected type naturally across different areas of the tray and let everything settle.
4. Enter or speak the actual upward-facing values from left to right.
5. Review every die and label before saving.
6. Continue until every face has useful, varied coverage.

Nearly identical resting positions may be skipped. Roll again or reposition every die so the model receives a genuinely different view.

Advanced — Mixed Dice Training

Use Mixed Dice Training when you want a complete polyhedral set to learn together. It is especially useful for a newly purchased set containing several die types.

1. Choose the die types and quantities in the set.
2. Roll the complete set naturally, leaving enough room for every die to settle separately.
3. Follow the highlighted review from left to right, confirming die types before face values.
4. Remove any blurred or unusable crop instead of saving it.
5. Continue until the progress view shows that each die type has enough useful examples.

The workflow tracks progress for each die type and can reduce the expected roll as parts of the set become complete. Same-type Natural Roll Training remains the simpler choice when only one die type needs attention.

Step 2 — Review and train

1. Open Training → Check Training Images.
2. Set aside images that are blurry, cut off, unreadable, poorly lit, incorrectly labeled, or near-duplicates.
3. Open Training → 2. Train Recognition Model.
4. Select Auto Configure and Start Training.
5. Let training finish, confirm the intended model is active, and test fresh physical rolls.

New images do not affect recognition until model training completes. If the current model cannot be updated safely, Joe's Real Rolls may restart the training workflow as a Full Rebuild.

The improvement loop

1. Test fresh rolls under normal play conditions.
2. Note values that are Unknown or repeatedly wrong.
3. Add a modest number of accurate, varied examples.
4. Review image quality and face balance.
5. Train again, confirm the active model, and retest.

If you have more than 100 careful physical examples for one die type and results remain poor, check lighting, focus, tray contrast, calibration, labels, and image quality before adding more. Starting that die type again in a clean, consistent setup can be more effective than extending a weak data set.

Back up a model that works

Open Maintenance → Backup / Restore Saved Models and create a clearly named backup before a major training change. You can restore a saved personal model or the protected Factory Baseline without deleting your training images or roll statistics.

Training troubleshooting

New images are not improving recognition

Saved images do not affect recognition until Training → 2. Train Recognition Model finishes. Confirm the intended Dice Set is active, complete training, verify which model became active, and then test fresh rolls rather than previously captured images.

Captures are skipped as repeats

Move and rotate every die, roll into another part of the tray, and allow the image to settle. Nearly identical resting positions add little useful variety and may be skipped intentionally.

Training is slow

• Let the current run finish; early progress does not represent the final model.
• Use Auto Configure and Start Training rather than guessing advanced settings.
• Close other processor- or graphics-heavy applications.
• Use the in-app progress view to follow the training run without interrupting it.
• CPU training is supported and can take longer than GPU-assisted training.

Training restarts as a Full Rebuild

If the active personal model cannot be updated safely, Joe's Real Rolls can restart with a Full Rebuild. Let the restarted run complete before judging the result. The protected Factory Baseline remains separate.

A recent training run made results worse

1. Restore your named known-good model backup.
2. Review recent images for wrong labels, glare, blur, cut-off dice, and duplicates.
3. Confirm the camera setup and lighting have not changed since capture.
4. Add a modest number of accurate, varied examples for the weak values.
5. Train the affected die type and retest with fresh rolls.

One face remains weak

Open the Recognition Improvement Guide and use Advanced Fixed Learning for the specific value. Preserve dots or underlines for 6/9 and 60/90, and vary position and rotation without hiding the orientation mark.

You already have more than 100 examples

Stop adding images and inspect the setup. Poor lighting, soft focus, weak tray contrast, incorrect labels, or inconsistent calibration can outweigh a large data set. Starting that die type again in a clean, stable setup can be more effective than extending weak data.