Phips/lucid-cc0-v2-hc

LUCID CC0 v2 HC — 256×256 High-Complexity SISR Finetuning Dataset

A high-complexity subset of LUCID CC0 v2 for finetuning SISR models on the most informative content. Contains only tiles with ICNet complexity ≥ 0.85, ensuring the model focuses on complex textures, edges, and patterns.

Format: WebDataset .tar shards (~1 GB each). Optimized for streaming training.

Statistics

Metric	Value
Tiles	193,262
Resolution	256×256 PNG
Total size	~27 GB
Shards	25 tar files (~1 GB each)
Source dataset	LUCID CC0 v2 (1.17M tiles)
ICNet complexity threshold	≥ 0.85
CLIP-IQA quality threshold	≥ 0.3
Mean ICNet complexity	0.917

Filtering Pipeline

This dataset applies two quality gates on top of the base LUCID CC0 v2 filtering (implemented in lucid-sisr):

1. ICNet complexity ≥ 0.85 — retains only the most complex tiles (high edge density, rich textures, fine detail). This is a stricter threshold than the base dataset (≥ 0.6).

2. CLIP-IQA quality ≥ 0.3 — removes ringing/haloring artifacts from aggressive sharpening.

The combination ensures every tile contains high-value, artifact-free content for efficient finetuning.

Purpose

High-complexity finetuning is the second stage of the three-stage pipeline:

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

After pretraining on the full dataset, finetuning on high-complexity content teaches the model to handle the most challenging reconstruction scenarios — fine textures, sharp edges, and intricate patterns.

Usage

Loading with WebDataset

import webdataset as wds

dataset = (
    wds.WebDataset("hf://datasets/Phips/lucid-cc0-v2-hc/shard-{00000..00024}.tar")
    .decode("pil")
    .to_tuple("png")
)
for image in dataset:
    # image is a PIL Image
    pass

Finetuning

python -m traiNNer.train -opt configs/train/HAT/HAT_M_LUCID_Finetune_HC.yml

Update pretrain_network_g in the config to point to your pretrained checkpoint from Stage 1. Use with traiNNer-redux and the HAT model.

Related Datasets

• Phips/lucid-cc0-v2 — Full dataset (256×256, 1.17M tiles). For Stage 1 pretraining.
• Phips/lucid-cc0-v2-hc-512 — High-complexity 512×512 subset (101K tiles). For Stage 3 finune-finetuning.

License

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

Citation

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

Links

• Filtering pipeline: Phhofm/lucid-sisr
• Source dataset: nyuuzyou/pxhere