darktable-org/Upscale-RealPLKSR-ONNX

RealPLKSR (ONNX)

ONNX conversion of RealPLKSR – the Real-world variant of Partial Large Kernel CNNs for Super-Resolution. Includes 2x and 4x scale factors, MSSIM-pretrain stage (no GAN finetune).

Packaged for use with darktable's neural restore module via the darktable-ai pipeline.

Files

Path	Purpose
onnx/model_x2.onnx	2x upscaler, static input 512×512 → output 1024×1024
onnx/model_x4.onnx	4x upscaler, static input 256×256 → output 1024×1024
checkpoints/2x_realplksr_mssim_pretrain.pth	Original PyTorch weights for the x2 model
checkpoints/4x_realplksr_mssim_pretrain.pth	Original PyTorch weights for the x4 model
config.json	HF model metadata

The checkpoints/ directory holds the original .pth files used to produce the ONNX exports. They are kept here so the conversion can be reproduced even if the upstream Google Drive links stop being reachable.

Source

• Architecture and weights: dslisleedh/PLKSR (MIT) – Dongheon Lee et al., Machine Intelligence Laboratory, University of Seoul
• Paper: Partial Large Kernel CNNs for Efficient Super-Resolution (2024)
• Checkpoint release thread: dslisleedh/PLKSR#4 ("Real-world PLKSR")
• Trained via the neosr framework with the RealESRGAN degradation pipeline
• Conversion code: darktable-org/darktable-ai under GPL-3.0+
• Original Google Drive links (subject to availability):
  • x2: https://drive.google.com/file/d/1GAdf5VOqYa5ntswT9sYsKKZ2Z7OQp7gO/view
  • x4: https://drive.google.com/file/d/12ek1vitEporWc5qqaYo6AMy0-RYlRqu8/view

Usage

Inference inputs are RGB images in the [0, 1] range. The graphs have static input dimensions, so callers must tile at exactly the declared size. See the darktable-ai conversion script for the full export configuration, and the demo script for an ONNX Runtime example that handles tiling, mirror-padding, and overlap stitching.

License

• Model weights (in onnx/ and checkpoints/): MIT, from upstream dslisleedh/PLKSR.
• Conversion script and packaging: GPL-3.0+, see darktable-org/darktable-ai.

Notes

• Training dataset is not documented by the weights author; assumed to be DF2K (DIV2K + Flickr2K) per neosr common practice. Flickr2K does not carry an explicit open-source license.
• MSSIM-pretrain checkpoints only (no GAN finetune) – conservative output, no hallucinated detail.