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	Active baseline	Traditional static HDR fusion from MIT5K DNG images.
AITM	Placeholder	Future AI tone mapping from linear/HDR images to display-ready images.
DiffIPE	Placeholder	Future diffusion-based image post-enhancement.

The implemented modules are staged deliberately: raw-sim feeds JDD, and the current HDR module provides a traditional static fusion baseline before AI-based HDR fusion is added later.

Pipeline Overview¶

The current training flow is:

sRGB image
  -> raw-sim camera degradation
  -> noisy packed RAW burst
  -> JDD restoration network
  -> clean linear RGB
  -> traditional static HDR fusion baseline
  -> future AI tone mapping

raw-sim is responsible for physically motivated degradation. JDD restores noisy RAW bursts to linear RGB. HDR currently keeps the traditional static pipeline simple so AITM can be brought up next.

Languages¶

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