---
license: cc-by-nc-4.0
pretty_name: PRISM
---

# PRISM

This dataset is released under the CC BY-NC 4.0 License.

## Dataset Structure

After extracting the archives, the dataset has the following layout:

```text
PRISM/
  train/
    input/
      <gt_id>_<variant_id>.png
    gt/
      <gt_id>.png
  test/
    input/
      <gt_id>_<variant_id>.png
    gt/
      <gt_id>.png
  real-world-image/
    image_001.png
    ...
```

Each image in `input/` is a moire-contaminated fabric image. Its clean target is stored in `gt/` and is matched by the prefix before the underscore.

For example:

```text
train/input/000517_04.png  ->  train/gt/000517.png
test/input/001174_09.png   ->  test/gt/001174.png
```

One clean ground-truth image may correspond to multiple synthesized moire variants.
The `real-world-image/` folder contains unpaired real-world moire images for qualitative evaluation.

## Split Statistics

```text
Split   Moire inputs   Clean GT images
train   14,587         1,783
test    1,463          180
total   16,050         1,963
```

The number of paired samples is the number of files in `input/`, because each input image has one aligned clean target.



## Download and Verification

The test split is provided as `PRISM_test.zip`. Real-world images are provided
as `real-world-image.zip`. The training split is provided as parts under
`train_chunks/` and must be concatenated into `PRISM_train.zip` before
extraction.

SHA-256 checksums for the test archive, the real-world archive, all training
parts, and the reconstructed training archive are provided in `sha256sum.txt`.

Verify the downloaded files:

```bash
sha256sum -c sha256sum.txt --ignore-missing
```

Reconstruct the training archive from the split files:

```bash
cat train_chunks/PRISM_train.zip.part-* > PRISM_train.zip
```

Verify the reconstructed archive:

```bash
sha256sum -c sha256sum.txt --ignore-missing
```

Then extract the archives:

```bash
unzip PRISM_train.zip
unzip PRISM_test.zip
unzip real-world-image.zip
```

## Usage Notes

- Images are stored as RGB PNG files.
- The train/test split is fixed and should be kept unchanged for reproducible evaluation.
- Evaluation should be performed on `test/input/` using the corresponding targets in `test/gt/`.
- The filename prefix before `_` determines the ground-truth image ID.

## Citation

If you use PRISM, please cite the corresponding paper:

```bibtex
@inproceedings{wei2026fabric,
  title     = {Fabric Image Demoir{\'e}ing Benchmark from Synthesis to Restoration},
  author    = {Wei, Pengchao and Guo, Xiaojie},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}
```