claritycore

[!!!WORK IN PROGRESS!!!]Easy and fast low-level vision for everyone

[!!!WORK IN PROGRESS!!!]Next-generation Open Source toolkit for low-level vision. Engineered for state-of-the-art performance in image and video pixel2pixel tasks, including Super-Resolution, Denoising, Deblurring, and more.

If you find ClarityCore useful, please consider giving it a star ⭐ and Join Discord!

Latest News 🚀

• I am very busy at my current position. I want to continue this project in ~January
• [2025/11] Major refactoring: unified dataset, preset system, Rich CLI
• [2025/10] Project start

🚀 Quick Start

Installation

pip install claritycore

Or install from source:

git clone https://github.com/Bezdarnost/claritycore.git
cd claritycore
pip install -e .

Training

Train a 4x super-resolution model with one command:

# List available presets
claritycore list

# Train with a preset
claritycore train rrdbnetx4

# Train with custom settings
claritycore train rrdbnetx4 --steps 100000 --batch-size 8

Python API

from claritycore import AutoConfig, AutoModel, Trainer
from claritycore.data import Pixel2PixelDataset, DatasetConfig

# Create model
config = AutoConfig.from_name("rrdbnet", scale=4)
model = AutoModel.from_config(config)

# Create dataset
data_config = DatasetConfig(target_dir="data/HR", scale=4)
dataset = Pixel2PixelDataset(data_config)

# Train
trainer = Trainer(model, train_loader, optimizer, training_config)
trainer.train()

📦 Available Models

Model	Task	Parameters	Scales
RRDBNet	Super-Resolution	16.7M (full) / 1M (lite)	2×, 3×, 4×, 8×

See docs/models.md for detailed documentation.

📁 Dataset Structure

ClarityCore expects datasets in this structure:

datasets/DIV2K/
├── hr/           # Target (high-resolution) images
├── x2/           # Input images for 2x SR
├── x3/           # Input images for 3x SR
└── x4/           # Input images for 4x SR

The Pixel2PixelDataset automatically:

• Detects filename suffix patterns (e.g., 0001x4.png)
• Generates input images on-the-fly if no input directory exists
• Supports flexible normalization ([0,1], [-1,1], or custom)

🛠️ CLI Reference

# Show help
claritycore --help

# List available training presets
claritycore list

# Train with preset
claritycore train <preset> [options]

# Examples
claritycore train rrdbnetx4                    # Full RRDBNet, 4x SR
claritycore train rrdbnet-litex4               # Lightweight variant
claritycore train rrdbnetx4 --steps 50000      # Custom steps
claritycore train rrdbnetx4 --batch-size 8     # Custom batch size

# Advanced training (full control)
claritycore train --model rrdbnet --scale 4 --data path/to/data

📊 Presets

Preset	Model	Scale	Config	Steps
rrdbnetx2	RRDBNet	2×	64 feat, 23 blocks	400K
rrdbnetx4	RRDBNet	4×	64 feat, 23 blocks	400K
rrdbnet-litex4	RRDBNet	4×	32 feat, 6 blocks	200K

Run claritycore list for all available presets.

📖 Citation

If you use ClarityCore in your research, please cite:

@software{ClarityCore_2025,
  author = {Urumbekov, Aman},
  license = {Apache-2.0},
  title = {{ClarityCore}},
  url = {https://github.com/Bezdarnost/claritycore},
  year = {2025}
}

Urumbekov, A. (2025). ClarityCore: A toolkit for low-level vision. GitHub. https://github.com/bezdarnost/claritycore

⭐ Star History

📄 License

Apache License 2.0. See LICENSE.