T2exture Checkpoints
This repository hosts the trained T2exture-S, T2exture-L, and T2exture-G checkpoints for source-conditioned thermal texture propagation.
Files
| File | Variant | Backbone |
|---|---|---|
t2exture-s.pt |
T2exture-S | AMT-S |
t2exture-l.pt |
T2exture-L | AMT-L |
t2exture-g.pt |
T2exture-G | AMT-G |
configs/train.yaml |
training config | five-frame centered passive context |
Each checkpoint stores the full T2exture model state for inference. A separate AMT initialization checkpoint is only needed when training from scratch.
Code And Data
Code release:
https://github.com/dccc2025/T2exture
Set the model and dataset repo ids locally:
export T2EXTURE_MODEL_REPO=<hf-model-repo-id>
export T2EXTURE_DATASET_REPO=<hf-dataset-repo-id>
Download the dataset:
python -c "import os; from huggingface_hub import snapshot_download; snapshot_download(repo_id=os.environ['T2EXTURE_DATASET_REPO'], repo_type='dataset', local_dir='datasets')"
Synthetic Inference
python -B infer.py --mode synthetic --variant l --data-root datasets --split test --output-dir outputs/infer/synthetic-l
infer.py downloads t2exture-<variant>.pt from T2EXTURE_MODEL_REPO when --checkpoint is not provided.
Real Inference
For real sequences, provide either residual texture anchors or source-on/source-off frames:
real/<sequence>/texture/001.png
real/<sequence>/passive/001.png
real/<sequence>/source_on/001.png
real/<sequence>/source_off/001.png
If source-off frames at active keyframes are missing, generate Stage 1 caches first:
python -B scripts/stage1.py --mode real --data-root datasets/real --real-config configs/real.yaml --backbone amt-l --pretrained pretrained/amt-l.pth --output-dir datasets/real/source_off/amt-l
Then run inference:
python -B infer.py --mode real --variant l --data-root datasets/real --source-off-root datasets/real/source_off/amt-l --output-dir outputs/infer/real-l
Notes
The released config uses a five-frame centered passive structural context and active stride 10. Evaluation numbers are intentionally kept with the paper and local experiment artifacts, not duplicated in this model card.