TerraDiT

TerraDiT / TerraDiT-Ω weights

Official weights for TerraDiT (α, Σ) and TerraDiT-Ω (ECCV 2026), diffusion transformers for satellite image synthesis at 256x256 conditioned on text, geolocation, point prompts, and arbitrary geospatial primitives.

Code · Collection · Data · TerraDiT-Ω paper · project page · TerraDiT paper · project page

Each folder holds model.safetensors (EMA weights, fp16, REPA projector heads removed) and config.json (family, architecture, conditioning flags, training steps, parameter count).

Usage

git clone --recursive https://github.com/mvrl/TerraDiT.git && cd TerraDiT && pip install -e .
python terradit/omega_demo.py            # downloads omega_xl on first use

from terradit.generation import build_inference_model
model = build_inference_model("omega", None, "omega_xl", "cuda")   # name -> auto-download

Sampling: Euler, 100 steps, no classifier-free guidance, SDXL VAE decoder. Text encoder: LongCLIP (zer0int/LongCLIP-KO-LITE-TypoAttack-Attn-ViT-L-14, 144 tokens). Geolocation: RANGE+ (1280-d). See the GitHub repo for demos, evaluation, and training.

Training data

Git-10M imagery (snapshot 29f192b8, 256x256, ~1 m/px) with derived OSM conditioning and RANGE+ embeddings from MVRL/TerraDiT-data.

License and attribution

Weights: CC-BY-NC-4.0 (non-commercial research use), because the training imagery, Git-10M, is CC-BY-NC-ND-4.0. Code: Apache 2.0. Conditioning data derives from OpenStreetMap: contains information from OpenStreetMap, made available under the Open Database License (ODbL), © OpenStreetMap contributors. Full component list: NOTICE.md in the code repository.

Intended use and limitations

Research on controllable remote-sensing image synthesis, data augmentation, and map-to-image translation. Outputs are synthetic and must not be presented as real imagery. The models inherit Git-10M's geographic and semantic coverage; OSM tag coverage is uneven across regions.

Citation

@inproceedings{wei2026terraditomega,
  title     = {TerraDiT-{\Omega}: Unified Spatial Control for Satellite Image Synthesis with Any Geospatial Primitive},
  author    = {Wei, Brian and Sastry, Srikumar and Cher, Daniel and Xing, Eric and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision}, year = {2026}
}
@article{sastry2026terradit,
  title   = {TerraDiT: Point-Conditioned Diffusion Transformer for Satellite Image Synthesis},
  author  = {Sastry, Srikumar and Cher, Daniel and Wei, Brian and Dhakal, Aayush and Khanal, Subash and Gupta, Dev and Jacobs, Nathan},
  journal = {arXiv:2603.02172}, year = {2026}
}