On March 13th, it was announced that the Xiaomi Mi 14 Ultra, hailed as the Leica Imaging King, was officially unveiled. This marks Xiaomi's most powerful imaging flagship to date, boasting not only a new Summilux All-Star quad-camera setup but also debuting the AI-powered computational photography platform, Xiaomi AISP.
Xiaomi AISP represents a significant upgrade to Xiaomi's imaging capabilities. Today, Xiaomi's official communication provided in-depth insights into the Xiaomi AISP platform.
Xiaomi has stated that the Xiaomi AISP platform utilized in the newly launched 14 Ultra is developed based on the Stable Diffusion model, which is a foundational AI large model in the field of image processing. This platform is the result of innovative thinking focused on efficacy, computational power, and generation.
Xiaomi has split Xiaomi AISP into four major models: Fusion Optical Model, Tone Model, Color Model, and Portrait Model. Each model has undergone acceleration processing such as miniaturization and heterogeneous parallelism to meet the objective requirements of smartphone photography.
For the first time, Xiaomi has applied AIGC technology to computational photography, bringing precise predictions of "what to see" and "how to present" images, addressing issues such as blur and lack of authenticity in long-distance shots with the camera's telephoto function.
Utilizing the Stable Diffusion mega-model, Xiaomi 14 Ultra introduces an AI-powered Ultra Zoom feature, transforming images into high-resolution marvels.
Thanks to Xiaomi AISP, the FusionLM algorithm revolutionizes the conventional approach by preserving raw optical information, ensuring linear calibration of light intensity, and seamlessly synthesizing images to overcome prior issues of unnatural transitions.
Moreover, to enhance realism and replicate human-eye perception, Xiaomi employs the ToneLM algorithm for tone mapping, achieving lifelike lighting effects and closer-to-reality light ratios. This model autonomously adjusts brightness and contrast on Ultra Raw data, surpassing traditional tone mapping algorithms in handling complex scenes.
Lastly, the Portrait Mega-Model encompasses three key modules: "portrait semantics," "portrait blur," and "high-fidelity portrait beautification." By mimicking the mindset of photography masters, this algorithmic approach enhances depth, realism, and skin texture, offering a natural portrayal of subjects.
