---
library_name: pytorch
pipeline_tag: image-to-image
tags:
- thermal-video-interpolation
- frame-interpolation
- t2exture
- pytorch
datasets:
- chenjiashuo/T2exture_datasets
---

# T2exture Checkpoints

This repository contains the trained T2exture-S, T2exture-L, and T2exture-G checkpoints.

## Files

| File | Model | Backbone |
| --- | --- | --- |
| `t2exture-s.pt` | T2exture-S | AMT-S |
| `t2exture-l.pt` | T2exture-L | AMT-L |
| `t2exture-g.pt` | T2exture-G | AMT-G |

Each file stores the full T2exture model state and can be used for inference without a separate AMT initialization checkpoint.

## Quick Start

Download the code, dataset, and one checkpoint into the documented local layout. The dataset download includes the matching Stage 1 source-off caches:

```bash
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='chenjiashuo/T2exture_datasets', repo_type='dataset', local_dir='datasets')"
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='chenjiashuo/T2exture_model', filename='t2exture-g.pt', local_dir='pretrained/t2exture_model')"
```

Run synthetic inference with the matching cache:

```bash
python -B infer.py \
  --mode synthetic \
  --variant g \
  --checkpoint pretrained/t2exture_model/t2exture-g.pt \
  --data-root datasets \
  --source-off-root datasets/source_off/amt-g \
  --split test \
  --output-dir outputs/infer/t2exture-g-test
```

For real inference, provide real frames and a matching Stage 1 source-off cache, then pass it with `--source-off-root`. The cache is required at active keyframes because the method constructs anchors as:

```text
X_k = [S^on_k - S_hat^off_k]_+
```

The public code README documents the full S/L/G training and evaluation commands. This model repository contains weights only; it does not contain training logs or experiment outputs.