neosr Framework

neosr is an open-source framework designed for training super-resolution models. It provides a reproducible environment for image restoration across various domains, including photos, anime/cartoons, illustrations, medical imaging, forensics, and geospatial data.

Key characteristics:

• Accessible: Supports a wide range of networks, losses, optimizers, and augmentations via configuration.
• Efficient: Optimized for faster training, quicker convergence, and low GPU requirements.
• Reproducible: Provides deterministic training environments for bit-exact reproducible models on the same platform.
• Simple: Uses readable Python code managed by uv, ruff, mypy, and torchfix.

The neosr project is released under the Apache 2.0 license.

Because neosr integrates various architectures, optimizers, losses, and augmentations from the research community, it includes numerous third-party components. These components are subject to their own respective licenses, including but not limited to:

• Apache 2.0 (e.g., BasicSR, SwinIR, OmniSR, DAT, DITN, etc.)
• MIT (e.g., HAT, Focal Frequency Loss, DiffJPEG, Real-CUGAN, etc.)
• BSD / BSD 3-Clause (e.g., Real-ESRGAN, patchgan)
• CC-NC 4.0 / CC-NC-SA 4.0 (e.g., SRFormer, DISTS weights, TOPIQ)

Users should refer to the specific license of each component if they are using specific architectures or modules within neosr for commercial or redistributed purposes.

To begin training a super-resolution model, use the neosr-train command followed by the path to your configuration file (typically a .toml file).

Example:

neosr-train config.toml

Configuration templates are available in the options/ directory of the repository. For a detailed explanation of available configuration options, refer to the Configuration Walkthrough in the project wiki.

Install neosr using the platform-specific installation scripts. Note that neosr will be installed in your terminal's current path.

Prerequisites:

• CUDA >= 12.4

Windows (PowerShell): Run the following command in PowerShell:

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/neosr-project/neosr/refs/heads/master/install_windows.ps1 | iex"

Linux (Bash): Run the following command in your terminal:

curl -LsSf https://raw.githubusercontent.com/neosr-project/neosr/refs/heads/master/install_linux.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/neosr-project/neosr/refs/heads/master/install_windows.ps1 | iex"

# OR for Linux
curl -LsSf https://raw.githubusercontent.com/neosr-project/neosr/refs/heads/master/install_linux.sh | sh

Specify validation metrics using the following options:

When training with GAN-based losses, specify the discriminator network using the following options:

Enable data augmentations during training using these options:

Choose the dataset loading strategy:

Select the high-level model type: