The University of Würzburg's reference darkroom

What it does

AISP is the University of Würzburg’s research collection for computational photography and low-level computer vision. It gathers reference implementations, training notebooks, and challenge baselines for NTIRE and AIM workshops, covering learned image signal processing, RAW restoration and super-resolution, RAW reconstruction from RGB, bokeh rendering, and lightweight smartphone enhancement. Most contributions are organized as Jupyter notebooks with end-to-end guides and synthetic data tutorials.

The interesting bit

The project treats the camera pipeline as a reversible learning problem: one branch reconstructs RAW sensor data from processed RGB images, while another learns to synthesize RGB from RAW using lightweight dictionaries or TFLite-deployable networks. The LPIENet work is explicitly optimized for mid-level smartphones, with the authors reporting 4K inference under one second and 50 times fewer MACs than NAFNet.

Key highlights

• Official baseline code for multiple NTIRE (CVPR) and AIM (ICCV/ECCV) challenge tracks, including RAW super-resolution, restoration, and reconstruction from sRGB.
• LPIENet ships in PyTorch and TensorFlow with TFLite conversion, targeting real-time perceptual enhancement on commercial smartphones.
• The AAAI 2022 Oral work on model-based ISPs uses learnable dictionaries for hybrid data-driven RAW reconstruction and augmentation.
• Includes reference implementations for blind RAW super-resolution (WACV 2024) and multi-lens bokeh rendering (CVPRW 2023).
• Provides tutorial notebooks for generating degraded RAW images with controlled noise, blur, and downsampling.

Caveats

• Some sections, such as the AAAI 2022 learnable dictionaries work, direct readers to contact the author for implementation support or qualitative samples rather than offering fully self-contained code.
• The repository is explicitly challenge-centric and paper-oriented, so expect a directory of disparate research artifacts rather than a unified API or stable library interface.

Verdict

A useful starting point if you compete in NTIRE or AIM challenges, or if you are prototyping RAW restoration and mobile ISP pipelines. Look elsewhere if you need a polished, product-ready imaging SDK.