Apple study shows how AI can improve low-light photos - 9to5Mac

Apple researchers, in collaboration with Purdue University, developed an AI model named DarkDiff to significantly enhance extremely dark, low-light raw images. Traditional methods often result in oversmoothing or loss of detail due to noise reduction. DarkDiff tackles this by retasking a pre-trained generative diffusion model, similar to Stable Diffusion, directly into the camera's Image Signal Processor (ISP) pipeline, rather than just post-processing. This allows the model to understand and recover fine details and accurate colors from raw sensor data by computing attention over localized image patches, mitigating hallucinations. While DarkDiff outperforms state-of-the-art methods in perceptual quality, the researchers note its high computational demands suggest it would likely require cloud processing rather than running locally on a device soon. The study highlights Apple's ongoing commitment to computational photography advancements.

(Source：9to5Mac)