Open-AISP¶

Open-AISP is an open-source AI-ISP playground for building an end-to-end image signal processing pipeline. It starts from synthetic raw degradation and extends toward neural reconstruction, HDR synthesis, tone mapping, and diffusion-based image enhancement.

The project is organized into five modules:

Module	Status	Purpose
raw-sim	Active	Converts sRGB images into noisy packed RAW and clean linear RGB supervision.
JDD	Active	Restores clean linear RGB from noisy RAW bursts with joint denoising and demosaicing.
HDR	Placeholder	Future multi-frame high dynamic range synthesis from exposure brackets.
AITM	Placeholder	Future AI tone mapping from linear/HDR images to display-ready images.
DiffIPE	Placeholder	Future diffusion-based image post-enhancement.

The two implemented modules are intentionally coupled through a single degradation interface: JDD calls raw-sim to produce RAW training samples. When the camera model, CFA layout, PSF blur, or noise model changes in raw-sim, the JDD data pipeline follows the same behavior.

Pipeline Overview¶

The current training flow is:

sRGB image
  -> raw-sim camera degradation
  -> noisy packed RAW burst
  -> JDD restoration network
  -> clean linear RGB

raw-sim is responsible for physically motivated degradation. JDD is responsible for burst restoration from noisy RAW to linear RGB.

Languages¶

English is the default documentation language. A Chinese version is available from the language selector.