Inference¶

Single-image DTM prediction¶

PYTHONPATH=src uv run python scripts/inference/inference.py \
    --ckpt /path/to/best.ckpt \
    --image /path/to/PSP_XXXXXX_XXXX_RED.JP2 \
    --output /path/to/predicted_dtm.tif

The script runs an Euler ODE integration of the flow-matching velocity field, with optional ensemble averaging over multiple stochastic starts for variance reduction.

Pre-computing VAE latents¶

For larger-scale batched inference, latents can be pre-computed:

PYTHONPATH=src uv run python scripts/inference/precompute_latents.py \
    --root /scratch/mars_hirise --output /scratch/latents

This script uses the legacy filesystem dataset at scripts/inference/dtm_dataset.py.

Tile-based reconstruction¶

For inference on full strips, predictions are emitted per-patch and then merged with Huber-IRLS overlap blending + cosine taper:

PYTHONPATH=src uv run python scripts/reconstruction/surface_blend.py \
    --pred-dir /path/to/per_patch_preds \
    --output /path/to/strip_dtm.tif

See scripts/inference/inference.py for the CLI surface and depth_fm.training.lightning_module for the underlying model wrapping.