Apppublic

hugging-apps/nasa-ibm-ai4science-nasa-ibm-lunar-foundation-model

sourceHugging Faceapache-2.0updated 16d agoView on Hugging Face

2likes

App README

NASA-IBM Lunar Foundation Model — any-to-any probe

Interactive demo of `nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model`, a TerraMind-style multimodal foundation model for the Moon (ViT-B encoder-decoder, 11 co-registered LRO modalities, per-modality VQ tokenizers with diffusion decoders).

The Space runs the model's any-to-any generation path, following the authors' reference notebook `examples/generate_images.ipynb` one-to-one (same config.yaml, same modality info and dataset transforms, same get_lfm_generation_model / load_ckpt_for_generation / prepare_model_inputs / denormalize_data calls, same default sampling parameters: 25 detokenizer steps, 1 ROAR decoding step, temperature 1.0, top-p 0.8).

Pick a sample tile, choose which modalities to condition on and which one to hold out, and the backbone predicts the held-out modality's tokens; its VQ detokenizer decodes them back to an image. Alongside the generation the demo shows the held-out original, the tokenizer's own reconstruction of that original (the achievable ceiling) and a signed difference map.

Scope and caveats

Straight from the model card: this is a qualitative probe of learned cross-modal structure, not a scientific-grade generative product. Absolute values drift, outputs carry no geodetic reference frame, and nothing here should be used for mission planning or quantitative science. Judge the spatial structure.

Data

Sample tiles are the authors' own SomBench pretraining sample (8 WAC regional tiles, 6 NAC local tiles), released under CC-BY-4.0 — credit to NASA IMPACT / IBM and the LRO WAC / NAC / LOLA source products. Multi-channel co-registered netCDF lunar tiles cannot be supplied by hand, so the demo works from these official samples rather than user uploads.

Code

ni_lfm/ is vendored from NASA-IMPACT/NASA-IBM-Lunar-Foundation-Model (Apache-2.0) so the Space can pin a ZeroGPU-compatible torch; see LICENSE and NOTICE for the full attribution chain (4M / MultiMAE, timm, diffusers, vector-quantize-pytorch, …).