Phips/lucid-cc0-v2-hc

LUCID-CC0 v2 High Complexity (256×256): Finetuning Dataset for SISR

A high-complexity subset of lucid-cc0-v2 containing only the most detailed, information-rich tiles. Designed for finetuning pretrained SISR models to push final quality metrics.

Overview

Property	Value
Source	lucid-cc0-v2 (filtered subset)
Filtering	ICNet complexity ≥ 0.85 (highest-detail tiles only)
Tile size	256×256 pixels
Mean complexity	~0.917
Total tiles	~271,000
Disk size	~35 GB
License	CC0-1.0 (public domain)

Intended Use

This dataset is designed for finetuning SISR models that have been pretrained on the full lucid-cc0-v2 dataset. The high-complexity tiles contain the sharpest edges, finest textures, and most intricate details — exactly what a model needs to refine its super-resolution capabilities.

Recommended Training Strategy

Stage	Dataset	Purpose
1. Pretrain from scratch	lucid-cc0-v2 (200GB)	Learn general image representations
2. Finetune	This dataset (27GB)	Refine on highest-quality tiles
3. Finetune-finetune	lucid-cc0-v2-hc-512 (512×512)	Push quality with max patch size

Why High Complexity?

• Transformer models (HAT, SwinIR) are data-hungry but also benefit from quality-focused finetuning
• High-complexity tiles contain more high-frequency information per pixel
• Finetuning on these tiles specifically improves texture recovery and edge sharpness
• The 256×256 patch size is compatible with most SISR training frameworks

Dataset Structure

lucid-cc0-v2-hc/
├── train/
│   ├── 000/
│   │   ├── 00000.png
│   │   └── ...
│   ├── 001/
│   └── ...
├── LR/
│   ├── x2/           # Bicubic downscaled ×2
│   └── x4/           # Bicubic downscaled ×4
└── metrics.csv       # Per-image complexity statistics

Filtering Criteria

Tiles were selected from lucid-cc0-v2 based on ICNet complexity score:

• Threshold: ≥ 0.85 (out of 1.0)
• Distribution: Mean 0.917, all tiles above 0.85
• What this captures: Images with strong edges, fine textures, intricate patterns, high local contrast
• What this excludes: Smooth gradients, blurry regions, low-detail areas, sky/water surfaces

Bicubic Downscaling

LR images are provided using MATLAB-compatible bicubic interpolation, matching the standard used in SISR benchmarks.

Scale factors: ×2 and ×4.

Citation

@dataset{lucid_cc0_v2_hc,
  title={LUCID-CC0 v2 High Complexity: Finetuning Dataset for SISR},
  author={Phips},
  year={2026},
  license={CC0-1.0},
  url={https://huggingface.co/datasets/Phips/lucid-cc0-v2-hc}
}