Hugging Face's logo

dccc2025
/
TeX-UNet

Image-to-Image
PyTorch
Safetensors
hyperspectral-imaging
long-wave-infrared
thermal-infrared
temperature-emissivity-texture
lwir
hadar
tex-1500
Eval Results (legacy)
Model card Files
xet
 Community
1
TeX-UNet / README.md
dccc2025's picture
dccc2025
Improve README with TeX-1500 details (#1)
3fa9ee8
preview code
|
raw
history blame contribute delete
10.1 kB
metadata 
license: apache-2.0
library_name: pytorch
pipeline_tag: image-to-image
tags:
  - hyperspectral-imaging
  - long-wave-infrared
  - thermal-infrared
  - temperature-emissivity-texture
  - lwir
  - hadar
  - tex-1500
  - arxiv:2606.03806
datasets:
  - jialelin2007/TeX-1500
metrics:
  - mean_absolute_error
  - mean_squared_error
  - spectral_angle_mapper
model-index:
  - name: TeX-UNet-v2 DARPA
    results:
      - task:
          type: image-to-image
          name: LWIR HSI to TeX decomposition
        dataset:
          name: TeX-1500 DARPA IH-test
          type: jialelin2007/TeX-1500
        metrics:
          - name: T MAE
            type: mean_absolute_error
            value: 7.3284
            unit: K
          - name: e MSE
            type: mean_squared_error
            value: 0.0453
          - name: X MSE
            type: mean_squared_error
            value: 0.0311
      - task:
          type: image-to-image
          name: LWIR HSI to TeX decomposition
        dataset:
          name: TeX-1500 FTIR-zeroshot-test
          type: jialelin2007/TeX-1500
        metrics:
          - name: T MAE
            type: mean_absolute_error
            value: 5.8309
            unit: K
          - name: e MSE
            type: mean_squared_error
            value: 0.0674
          - name: X MSE
            type: mean_squared_error
            value: 0.0219
  - name: TeX-UNet-v2 FTIR Few-shot
    results:
      - task:
          type: image-to-image
          name: LWIR HSI to TeX decomposition
        dataset:
          name: TeX-1500 FTIR-fewshot-test
          type: jialelin2007/TeX-1500
        metrics:
          - name: T MAE
            type: mean_absolute_error
            value: 4.1004
            unit: K
          - name: e MSE
            type: mean_squared_error
            value: 0.0458
          - name: X MSE
            type: mean_squared_error
            value: 0.022

TeX-UNet

TeX-UNet is the initial wavelength-aware neural baseline for TeX-1500, a paired real-world LWIR hyperspectral benchmark for temperature-emissivity-texture decomposition. Given a calibrated LWIR HSI cube and its wavelength positions, TeX-UNet predicts temperature T, emissivity e, and texture X.

This repository releases two inference checkpoints from TeX-1500: A Paired Real-World LWIR Hyperspectral Dataset and Benchmark for Temperature-Emissivity-Texture Decomposition.

Authors

*Equal Contribution.

Links

Files 

TeX-UNet/
  README.md
  LICENSE
  CITATION.cff
  NOTICE
  model_index.json
  verification.json
  requirements.txt
  docs/reproducibility.md

  tex_unet_v2_darpa/
    model.safetensors
    config.json
    normalization.json
    inference_config.yaml
    training_config.json
    metrics.json
    conversion.json

  tex_unet_v2_ftir_fewshot/
    model.safetensors
    config.json
    normalization.json
    inference_config.yaml
    training_config.json
    metrics.json
    conversion.json

Checkpoints

Directory Training data Intended use Parameters
tex_unet_v2_darpa DARPA IH training split DARPA IH test and FTIR zero-shot transfer 34.28M
tex_unet_v2_ftir_fewshot FTIR few-shot split, initialized from tex_unet_v2_darpa FTIR few-shot transfer 34.28M

Both checkpoints are stored as safetensors and contain only the model state dict, not optimizer state.

Each checkpoint uses the same TeX-UNet-v2 architecture:

  • input HSI tensor: [B, 64, H, W]
  • input wavelength tensor: [64] or [B, 64], in micrometers
  • outputs:
    • T: [B, 1, H, W], temperature
    • e: [B, 64, H, W], spectral emissivity
    • X: [B, 1, H, W], texture
  • wavelength range: 6-14 micrometers
  • inference tiling: 224 x 224 patches, stride 112, raised-cosine blending
  • released parameter count: 34,279,155

Benchmark

The initial benchmark reports TeX-UNet inversion results on held-out DARPA IH pushbroom scenes and FTIR zero-/few-shot transfer scenes.

Test split Checkpoint T MAE (K) T MAPE (%) e MSE e SAM (rad) X MSE X SAM (rad)
DARPA IH-test tex_unet_v2_darpa 7.3284 2.5488 0.0453 0.2267 0.0311 0.5206
FTIR-zeroshot-test tex_unet_v2_darpa 5.8309 1.9753 0.0674 0.0451 0.0219 0.2995
FTIR-fewshot-test tex_unet_v2_ftir_fewshot 4.1004 1.3830 0.0458 0.1970 0.0220 0.2224

e and X are normalized. Full training settings, inference tiling, and evaluation details are in the paper and in each checkpoint directory's training_config.json, normalization.json, and inference_config.yaml.

Use With The Code Repository

Install the released inference code from GitHub:

git clone https://github.com/dccc2025/TeX-1500
cd TeX-1500
uv sync

Download one checkpoint:

uv run python scripts/download_weights.py \
  --repo-id dccc2025/TeX-UNet \
  --filename tex_unet_v2_darpa/model.safetensors \
  --output checkpoints/tex_unet_v2_darpa.safetensors

Run inference with a calibrated HSI .mat file:

uv run tex1500-infer \
  --checkpoint checkpoints/tex_unet_v2_darpa.safetensors \
  --model-config configs/tex_unet_v2_model.json \
  --input /path/to/hsi.mat \
  --output outputs/sample_tex_prediction.npz

The input .mat should contain an HSI cube and wavelength positions compatible with the TeX-1500 data format. See the GitHub repository for the exact CLI arguments and output format.

Loading The Weights 

from pathlib import Path
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

weights = hf_hub_download(
    repo_id="dccc2025/TeX-UNet",
    filename="tex_unet_v2_darpa/model.safetensors",
)
state_dict = load_file(weights, device="cpu")
print(len(state_dict))

To verify a checkpoint against the TeX-1500 codebase:

python scripts/verify_load.py \
  --code-root /path/to/TeX-1500 \
  --model-dir tex_unet_v2_darpa

Data

The dataset is hosted separately at jialelin2007/TeX-1500. The local dataset release notes describe TeX-1500 as a paired LWIR HSI benchmark with 1,522 calibrated real-scene HSI-TeX pairs from DARPA pushbroom data and FTIR acquisitions.

The dataset repository is separate from this model repository. The current dataset release is a preview containing one example sample for format inspection and visualization; the full dataset is planned for a later release.

Current preview layout:

data/sample_0001/
  hsi.mat
  hsi_noisy.mat
  T.mat
  e.mat
  X.mat
  previews/
    T.png
    e.png
    X.png
    hsi_band.png

The preview .mat files use MATLAB v5 format:

File Variable Shape Dtype Description
hsi.mat denoised_hsi_original [260, 1500, 256] single Denoised calibrated LWIR HSI cube.
hsi.mat hsi_wav [1, 256] double HSI wavelength grid.
hsi.mat good_band_indices [1, 230] int64 0-indexed valid band indices used by the processing pipeline.
hsi.mat calibrated_sky [1, 256] double Calibrated sky signal estimate.
hsi.mat observed_sky [1, 256] double Observed sky signal estimate.
hsi.mat working_wav [1, 256] single Calibrated working wavelength grid.
hsi_noisy.mat hsi_noisy [260, 1500, 256] single Noisy input HSI cube.
T.mat T [260, 1500] single Temperature field.
e.mat e [260, 1500, 256] single Spectral emissivity field.
X.mat X [260, 1500] single Texture field.

Preview PNG files are for visual inspection only and should not be treated as numeric labels.

If the Hugging Face dataset repository requires authentication, log in and accept its access terms before downloading files:

hf auth login
hf download jialelin2007/TeX-1500 \
  data/sample_0001/hsi.mat data/sample_0001/T.mat \
  --repo-type dataset \
  --local-dir data/tex1500_sample

The dataset license and access terms are governed by the dataset repository.

License

This model repository is released under Apache-2.0. The TeX-1500 dataset files are hosted separately and are distributed under the PolyForm Noncommercial License 1.0.0 in the dataset repository. TeX-1500 dataset files are not included in this model repository.

Limitations

  • TeX-UNet is an initial supervised baseline, not a complete physical inverse solver.
  • Predictions depend on calibrated LWIR HSI inputs and valid wavelength positions; unsupported sensors or uncalibrated radiance can degrade results.
  • The released checkpoints target TeX-1500-style normalized e and X outputs.
  • The benchmark is reported on the paper's held-out DARPA IH and FTIR splits; use caution when interpreting results outside these settings.

Citation 

@misc{dai2026tex1500pairedrealworldlwir,
      title={TeX-1500: A Paired Real-World LWIR Hyperspectral Dataset and Benchmark for Temperature-Emissivity-Texture Decomposition},
      author={Cheng Dai and Jiale Lin and Hongyi Xu and Bingxuan Song and Ziyang Xie and Fanglin Bao},
      year={2026},
      eprint={2606.03806},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.03806},
}