Saltar para o conteúdo principal

Training Custom AI Models for Heightmap Upscaling

· Leitura de 2 minutos
Heightcraft Team
Heightcraft Team
Core Developers

In this post, we'll dive into one of Heightcraft's most powerful features: Training custom AI upscaling models.

While Heightcraft comes with a general-purpose model, training on your specific type of terrain (e.g., rocky mountains, sand dunes, or urban landscapes) can yield significantly better results.

Why Train a Custom Model?​

The default model is trained on a diverse dataset of DEMs (Digital Elevation Models). However, terrain features vary wildly. A model trained on Swiss Alps data might not perform well on Martian crater data.

By training a custom model, you teach the AI to hallucinate details specific to your target biome.

Step 1: Prepare Your Dataset​

You need a set of high-resolution heightmaps. These will be the "ground truth".

  • Format: .tiff or .png (16-bit or 32-bit preferred).
  • Structure: Place them all in a single directory.
mkdir my_dataset
cp /path/to/high_res_tiffs/*.tiff my_dataset/

Step 2: Run Training​

Use the --train flag to start the training process. Heightcraft handles data augmentation and splitting automatically.

heightcraft --train \
  --dataset_path ./my_dataset \
  --epochs 50 \
  --batch_size 16 \
  --learning_rate 0.0001 \
  --pretrained_model my_custom_model.h5

Arguments:

  • --dataset_path: Path to a folder containing high-res images (PNG, JPG, TIF).
  • --pretrained_model: Path to save the trained model (default: trained_model.h5).

Hyperparameters:

  • --epochs (Default: 10): How many times the AI sees the entire dataset.
    • Effect: Higher values (e.g., 50-100) generally lead to better quality but take longer. If set too high, the model might "memorize" the training data (overfitting) and perform poorly on new maps.
  • --batch_size (Default: 16): How many images are processed at once.
    • Effect: Larger batches (32, 64) speed up training but require more GPU memory. Smaller batches (4, 8) are slower but offer more stable updates.
  • --learning_rate (Default: 0.0001): How fast the AI adapts its internal weights.
    • Effect: A higher rate (e.g., 0.001) learns faster but might miss the optimal solution (unstable). A lower rate (e.g., 0.00001) is more precise but takes much longer to converge.

Step 3: Upscale with Your Model​

Once training is complete, you can use your new model to upscale low-resolution inputs.

heightcraft low_res_input.png \
  --upscale \
  --pretrained_model my_custom_model.h5 \
  --upscale_factor 4 \
  --output_path high_res_output.png

Results​

You should see sharper edges and more plausible details compared to bicubic interpolation or the generic model.

Happy upscaling! 🚀