Phips/lucid-cc0-v2-hc-512

LUCID CC0 v2 HC 512 — 512×512 High-Complexity SISR Finetuning Dataset

The final stage of the LUCID three-stage training pipeline. Contains 512×512 high-complexity tiles for finetune-finetuning SISR models on high-resolution details. This is the highest-quality subset of the LUCID dataset family.

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

Statistics

Metric	Value
Tiles	100,866
Resolution	512×512 PNG
Total size	~51 GB
Shards	51 tar files (~1 GB each)
Source dataset	nyuuzyou/pxhere (CC0)
ICNet complexity threshold	≥ 0.85
CLIP-IQA quality threshold	≥ 0.3
Extraction scale	512×512
Filtering speed	~307 t/s (RTX 3060)

Filtering Pipeline

Tiles go through the same three-stage quality gate as the base dataset (implemented in lucid-sisr), but extracted at 512×512 resolution:

1. Signal filtering — removes flat/uninformative regions (entropy, Laplacian, gradient, blockiness)
2. ICNet complexity scoring ≥ 0.85 — ensures complex content (high edge density, rich textures)
3. CLIP-IQA quality filtering ≥ 0.3 — removes ringing/haloring artifacts
4. Deduplication — cosine similarity 0.96 removes redundant tiles

Note: ICNet scores 512×512 tiles lower than 256×256 tiles, so the threshold is set to 0.85 to maintain equivalent selectivity.

Purpose

This is the third and final stage of the training pipeline:

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

The 512×512 resolution allows the model to learn high-frequency details and fine textures that are lost at 256×256. This final finetuning stage produces the best visual quality for real-world super-resolution.

Usage

Loading with WebDataset

import webdataset as wds
from huggingface_hub import hf_hub_download
import glob, os

# Download shards (or use wds.WebDataset with HF URL)
dataset = (
    wds.WebDataset("hf://datasets/Phips/lucid-cc0-v2-hc-512/shard-{00000..00050}.tar")
    .decode("pil")
    .to_tuple("png")
)
for image in dataset:
    # image is a PIL Image
    pass

Finetune-finetune from a finetuned checkpoint

python -m traiNNer.train -opt configs/train/HAT/HAT_M_LUCID_FinetuneFinetune_HC_512.yml

Update pretrain_network_g in the config to point to your finetuned checkpoint from Stage 2. Use with traiNNer-redux and the HAT model.

Important: LR images should be created by the training software (traiNNer-redux) using MATLAB-compatible bicubic downscaling (a=-0.5 kernel), not pre-generated. This ensures fair comparison with other methods.

Related Datasets

• Phips/lucid-cc0-v2 — Full dataset (256×256, 1.17M tiles). For Stage 1 pretraining.
• Phips/lucid-cc0-v2-hc — High-complexity 256×256 subset (193K tiles). For Stage 2 finetuning.

License

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

Citation

@dataset{lucid_cc0_v2_hc_512,
  title={LUCID CC0 v2 HC 512: High-Complexity 512px SISR Finetuning Dataset},
  author={Phhips},
  year={2026},
  license={CC0-1.0},
  url={https://huggingface.co/datasets/Phips/lucid-cc0-v2-hc-512}
}

Links

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