EO-VAE

Earth Observation Variational Autoencoder. Self-contained, no external eo-vae package—uses only torch, safetensors, and huggingface_hub.

Usage

from huggingface_hub import snapshot_download
import sys
import torch

model_dir = snapshot_download("BiliSakura/EO-VAE")
sys.path.insert(0, model_dir)
from eo_vae import EOVAEModel, WAVELENGTHS

vae = EOVAEModel.from_pretrained(model_dir, torch_dtype=torch.float32, device="cpu")

x = torch.randn(1, 3, 256, 256)
wvs = torch.tensor(WAVELENGTHS["S2RGB"], dtype=torch.float32)
with torch.no_grad():
    recon = vae.reconstruct(x, wvs)
    z = vae.encode_spatial_normalized(x, wvs)  # [1, 32, 32, 32]

Model

Property	Value
Architecture	Flux-style VAE with wavelength-conditioned dynamic convolutions
Input	[B, C, 256, 256] + wavelengths [C]
Latent	[B, 32, 32, 32] (spatial) or [B, 128, 16, 16] (packed)
Modalities	S2RGB, S2L2A, S1RTC, S1GRD

Wavelengths (µm)

Modality	Values
S2RGB	0.665, 0.56, 0.49
S1RTC	5.4, 5.6
S2L2A	0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 1.610, 2.190, 0.945
S2L1C	0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.945, 1.375, 1.610, 2.190

Citation

@article{eo-vae,
  title={EO-VAE: Towards A Multi-sensor Tokenizer for Earth Observation Data},
  author={Lehmann, Nils and Wang, Yi and Xiong, Zhitong and Zhu, Xiaoxiang},
  journal={arXiv preprint arXiv:2602.12177},
  year={2026}
}