---
license: cc-by-nc-sa-4.0
library_name: pytorch
tags:
  - image-fusion
  - infrared-visible
  - perceptual-quality
  - pairwise-preference
  - bradley-terry
pipeline_tag: image-classification
---

# LPIFM — Learned Perceptual Image Fusion Measure

Pairwise perceptual preference model for **infrared–visible image fusion (IVIF)** ranking.

Given an IR source, a VI source, and two fused candidates, LPIFM predicts **A better**, **B better**, or **Tie**. Pairwise decisions can be aggregated with tie-aware Bradley–Terry (T-BT) to rank a method pool.

Code and examples: [github.com/HaoranLiu507/LPIFM](https://github.com/HaoranLiu507/LPIFM)

## Files

| File | Role |
| --- | --- |
| `lpifm_vifb_baseline_v1.pt` | Main VIFB-trained public checkpoint |
| `inference_config.yaml` | Decode defaults (`t`, `T_cal`, image size) |

EVAFusion fine-tuned weights are **not** hosted here; see the Zenodo companion archive when published.

**Note on checkpoint size.**
`lpifm_vifb_baseline_v1.pt` is about **2 GB** because it is a full training checkpoint: three weight copies (`swa` / `ema` / `model`, about 0.4 GB each) plus the optimizer (about 0.8 GB). To **use LPIFM** only, keep `swa_state_dict` (+ `config`); about 0.4 GB is enough. Inference already loads SWA by default.

## Architecture

- Backbone: **ConvNeXt-V2** (`convnextv2_base.fcmae_ft_in22k_in1k_384` via `timm`)
- Input size: **384 × 384**
- Task: source-conditioned pairwise preference scoring (ternary A / B / Tie)

## Decode / inference defaults

| Parameter | Value |
| --- | --- |
| Tie threshold `t` | `0.3` |
| Calibration temperature `T_cal` | `1.0` |
| Image size | `384` |

```text
d_cal = d / T_cal
d_cal >  0.3  → A better (0)
d_cal < -0.3  → B better (1)
otherwise     → Tie (2)
```

## Quick start

```bash
# from the GitHub repository
python scripts/download_assets.py --source hf
python predict.py \
  --config configs/release_inference.yaml \
  --dataset_root Dataset/VIFB \
  --ckpt checkpoints/lpifm_vifb_baseline_v1.pt \
  --image_name carLight.jpg \
  --a_dir U2Fusion \
  --b_dir SeAFusion
```

Or download this file directly:

```bash
huggingface-cli download FengShaner/LPIFM lpifm_vifb_baseline_v1.pt \
  --local-dir checkpoints
```

## Intended use

- Pairwise **perceptual preference** for **IVIF method ranking**
- Research / offline evaluation under the LPIFM protocol
- Not a general-purpose IQA model for arbitrary natural-image aesthetics

## Limitations

- Trained and validated under a specific preference-collection and ranking protocol
- Protocol mismatch (new dataset, different rater instructions, different method pools) may reduce agreement; fine-tuning may be required
- Non-commercial weights license (see below)

## Licenses

| Artifact | License |
| --- | --- |
| Model weights on this Hub repo | [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) |
| Source code on GitHub | [AGPL-3.0](https://github.com/HaoranLiu507/LPIFM/blob/main/LICENSE) |

Weights do **not** inherit AGPL; code does **not** inherit CC BY-NC-SA.

## Links

- GitHub: https://github.com/HaoranLiu507/LPIFM
- Hugging Face: https://huggingface.co/FengShaner/LPIFM
- Zenodo archival DOI: to be added after deposit publication

## Citation

See [`CITATION.cff`](https://github.com/HaoranLiu507/LPIFM/blob/main/CITATION.cff) in the GitHub repository.