geo-incontext-nf (f1024)

These are the weights of an in-context dense-map generator. Given 10 context (image, map) pairs and a query image, it generates the query's map. One model covers:

• segmentation;
• depth;
• remote-sensing regression from RGB, Sentinel-1 or Sentinel-2 inputs.

It is an autoregressive normalizing flow (STARFlow-style deep–shallow) on the 32×32×4 SD-VAE latent of the rendered map. The deep flow is a 36-layer dual-stream joint-attention transformer.

Code, installation, Docker, data layout and usage: https://github.com/Yusin2Chen/geo-incontext-nf

Files

file	content
geo_incontext_nf_f1024.safetensors	model weights, fp32, 2.75 B parameters (the frozen SD-VAE is loaded separately from stabilityai/sd-vae-ft-ema)
f1024.yaml	config used to build the model

Quick start

git clone https://github.com/Yusin2Chen/geo-incontext-nf && cd geo-incontext-nf
pip install -r requirements.txt && pip install -e .
python scripts/sample.py --config configs/f1024.yaml --data_root /path/to/raw_data \
    --weights Yusin/geo-incontext-nf --datasets openearthmap,nyuv2 --out outputs/samples

Training

steps	31.7 k
effective batch	128
hardware	4 × GH200
validation NLL (macro over datasets, step 30 k)	−0.250 nats/dim (seg −0.500, reg −0.136)

License

Research use only. The weights inherit the terms of their upstream components:

• the backbone was initialised from Ming-VideoMAR weights;
• the frozen SD-VAE is stabilityai/sd-vae-ft-ema;
• the training datasets each have their own license.