TerraMind-1.0-Tokenizer-S2L2A

This repository provides the S2L2A tokenizer checkpoint from TerraMind 1.0.

Model details

• Modality: S2L2A
• Input channels: 12
• Image size: 256x256
• Tokenizer backbone: vit_b_enc
• Quantization: FSQ (codebook_size=8-8-8-6-5, latent_dim=5)

Quick use (diffusers-style)

The tokenizer uses native diffusers patterns: ModelMixin, ConfigMixin, from_pretrained, and from_config.

from huggingface_hub import snapshot_download
import torch
import sys

# Download model repository
model_dir = snapshot_download("BiliSakura/TerraMind-1.0-Tokenizer-S2L2A")

# Expose local module, then load (diffusers-style)
sys.path.insert(0, model_dir)
from terramind_tokenizer import TerraMindTokenizer

# Load from path or Hub ID
tokenizer = TerraMindTokenizer.from_pretrained(
    model_dir,  # or "BiliSakura/TerraMind-1.0-Tokenizer-S2L2A"
    torch_dtype=torch.float32,
    device="cpu",
)

# S2L2A input: [B, 12, 256, 256]
x = torch.randn(1, 12, 256, 256)
tokens = tokenizer.tokenize(x)
print(tokens.shape)  # [1, 16, 16]

# Encode returns (quant, code_loss, tokens)
quant, code_loss, tokens = tokenizer.encode(x)

Load via AutoModel or TerraMindTokenizer (trust_remote_code)

You can load via diffusers AutoModel or the specific TerraMindTokenizer class with trust_remote_code=True:

from diffusers import AutoModel
import torch

# Option 1: AutoModel (auto-detects from config)
tokenizer = AutoModel.from_pretrained(
    "BiliSakura/TerraMind-1.0-Tokenizer-S2L2A",
    trust_remote_code=True,
    torch_dtype=torch.float32,
    device="cpu",
)

# Option 2: TerraMindTokenizer (explicit class)
from terramind_tokenizer import TerraMindTokenizer
tokenizer = TerraMindTokenizer.from_pretrained(
    "BiliSakura/TerraMind-1.0-Tokenizer-S2L2A",
    torch_dtype=torch.float32,
    device="cpu",
)

# Same API: tokenize(), encode()
x = torch.randn(1, 12, 256, 256)
tokens = tokenizer.tokenize(x)

Security: trust_remote_code=True runs code from the Hub. Only use with repos you trust. For production, pin a specific revision: revision="abc123def456" (commit hash after your changes).

Notes

• Uses diffusers ModelMixin and ConfigMixin for standard from_pretrained / from_config / save_pretrained.
• Supports both model.safetensors and diffusion_pytorch_model.safetensors weights.
• Tokenizer-focused API: tokenize(), encode(), and forward().
• Please follow TerraMind and TerraTorch licenses/usage terms from the upstream project.

Credits

• Original TerraMind project: https://github.com/IBM/terramind
• Original TerraMind model code (TerraTorch): https://github.com/terrastackai/terratorch/tree/main/terratorch/models/backbones/terramind
• This repository adapts tokenizer checkpoints for convenient Hugging Face usage.

Citation

If you use TerraMind in your research, please cite:

@article{jakubik2025terramind,
  title={TerraMind: Large-Scale Generative Multimodality for Earth Observation},
  author={Jakubik, Johannes and Yang, Felix and Blumenstiel, Benedikt and Scheurer, Erik and Sedona, Rocco and Maurogiovanni, Stefano and Bosmans, Jente and Dionelis, Nikolaos and Marsocci, Valerio and Kopp, Niklas and others},
  journal={IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025}
}