DEMForge POC

A practical proof-of-concept for DEM-trained procedural terrain generation.

Goal:

real DEM corpus
  -> paired DEM tiles
  -> train residual terrain detailer
  -> generate DEM-like fictional terrain
  -> render / QA / export

This POC intentionally starts with a conditional residual U-Net, not a giant diffusion model. Diffusion is the next spicy step after the boring model beats bicubic upscaling plus fractal nonsense.

Why this architecture first?

The first useful model should answer:

Given a coarse macro terrain heightmap, can we add believable real-world DEM-style detail?

Training target:

input  = coarse/upscaled terrain + derivative channels
target = high-res terrain residual
output = predicted residual
final  = coarse + predicted residual

This gives us control, low VRAM use, objective metrics, and a direct Minecraft terrain pipeline.

Quick smoke test

This does not need real DEM data. It generates fake DEM-ish tiles so you can test the training loop.

python scripts/make_synthetic_dataset.py --out data/synthetic --count 256 --size 256
python scripts/train_residual_unet.py --config configs/residual_unet_smoke.yaml
python scripts/sample.py --checkpoint outputs/checkpoints/best.pt --data data/synthetic/val --out outputs/samples
python scripts/render_heightmap.py --input outputs/samples/sample_000_pred.npy --out outputs/renders/sample_000_pred.png

Real DEM tile build

Put GeoTIFF DEM files under:

data/raw/

Then:

python scripts/build_tiles.py --src data/raw --out data/tiles --tile-size 512 --stride 256 --downscale 8

Expected split:

data/tiles/train/*.npz
data/tiles/val/*.npz

Output file format

Each tile .npz contains:

• x: model input, shape [C,H,W]
• y: residual target, shape [1,H,W]
• target: normalized high-res height, shape [1,H,W]
• coarse: normalized coarse/upscaled height, shape [1,H,W]
• meta: JSON string metadata

Notes

• Use real region-level validation splits. Do not randomly split neighboring tiles from the same mountain into train and val unless you want your validation score to lie like a used car salesman.
• Keep absolute elevation metadata, but train mostly on normalized local shape.
• Start 256x256 for smoke tests, then 512x512.