---
license: cc0-1.0
language:
- en
tags:
- super-resolution
- image-upscaling
- sisr
- training-dataset
- CC0
- pxhere
datasets:
- nyuuzyou/pxhere
metrics:
- psnr
- ssim
pipeline_tag: image-to-image
---

# LUCID CC0 v2 — 256×256 SISR Training Dataset

A large-scale, high-quality dataset for training single image super-resolution (SISR) models. Filtered from [nyuuzyou/pxhere](https://huggingface.co/datasets/nyuuzyou/pxhere) (CC0-licensed photography) using the [LUCID filtering pipeline](https://github.com/Phhofm/lucid-sisr).

## Statistics

| Metric | Value |
|--------|-------|
| Tiles | 1,169,792 |
| Resolution | 256×256 PNG |
| Total size | ~148 GB |
| Source images | ~33,000 |
| Source dataset | nyuuzyou/pxhere (CC0) |
| ICNet complexity threshold | ≥ 0.6 |
| CLIP-IQA quality threshold | ≥ 0.3 |
| Filtering speed | ~307 t/s (RTX 3060) |

## Filtering Pipeline

Every tile passes a **three-stage quality gate**:

1. **Signal filtering** — entropy, Laplacian variance, gradient energy, blockiness, noise ratio, aliasing ratio. Removes flat/uninformative regions (sky, water, solid colors).

2. **ICNet complexity scoring** — neural network measures local image complexity. Tiles below 0.6 threshold are removed. Ensures the model learns from informative content.

3. **CLIP-IQA quality filtering** — CLIP-based quality assessment detects ringing/haloring artifacts from aggressive sharpening. Tiles below 0.3 threshold are removed. This is critical because SISR models faithfully reproduce ringing artifacts as "details."

Key insight: **blur and noise don't hurt SISR** (the LR is just a bicubic downscaled HR), but **ringing IS problematic** — the model learns to reproduce halos as features. CLIP-IQA is the best metric for this (173.7% separation between ringing and clean images, 3.7× better than next best).

## Source

Images are sourced from [PxHere](https://pxhere.com/) under CC0 license — professional photography with permissive licensing. PxHere was chosen over other CC0 sources because:
- Pro photography (less social media sharpening bias)
- Diverse content (architecture, nature, portraits, objects)
- Clean originals (less aggressive post-processing)

## Usage

### Training from scratch

This dataset is designed for the **first stage** of a three-stage training pipeline:

```
Stage 1: Pretrain on this dataset (256×256, 1.17M tiles)
    ↓
Stage 2: Finetune on LUCID CC0 v2 HC (256×256, 193K tiles)
    ↓
Stage 3: Finetune-finetune on LUCID CC0 v2 HC 512 (512×512, 101K tiles)
```

Use with [traiNNer-redux](https://github.com/the-database/traiNNer-redux) and the HAT model:

```bash
python -m traiNNer.train -opt configs/train/HAT/HAT_M_LUCID_Pretrain.yml
```

### Dataset structure

```
train/
├── 000/   (≤10,000 PNG tiles per subdirectory)
├── 001/
├── 002/
└── ...
```

## Related Datasets

- **[Phips/lucid-cc0-v2-hc](https://huggingface.co/datasets/Phips/lucid-cc0-v2-hc)** — High-complexity subset (256×256, 193K tiles). For Stage 2 finetuning.
- **[Phips/lucid-cc0-v2-hc-512](https://huggingface.co/datasets/Phips/lucid-cc0-v2-hc-512)** — High-complexity subset (512×512, 18K tiles). For Stage 3 finetune-finetuning.

## License

CC0-1.0 (public domain). Source: PxHere.

## Citation

```bibtex
@dataset{lucid_cc0_v2,
  title={LUCID CC0 v2: Filtered SISR Training Dataset},
  author={Phhips},
  year={2026},
  license={CC0-1.0},
  url={https://huggingface.co/datasets/Phips/lucid-cc0-v2}
}
```