Hailuo 2.3 is a high-performance generative video model developed by MiniMax. Released as an iteration of the company's video-01 series, it is designed to synthesize high-fidelity cinematic video from both text and image prompts. The model utilizes a Noise-aware Compute Redistribution (NCR) framework, which optimizes the diffusion process by dynamically allocating computational resources, significantly improving training and inference efficiency compared to previous versions.
The model generates video at resolutions up to 1080p with durations typically ranging from 6 to 10 seconds. Version 2.3 introduces substantial improvements in temporal consistency, micro facial expressions, and complex physical interactions. It is particularly noted for its ability to render fluid human movements, such as dancing and gymnastics, while maintaining realistic physics for hair, clothing, and environmental elements like water and light.
Hailuo 2.3 supports a diverse range of visual aesthetics, including photorealism, anime, and stylized 3D animation. The model is available in two variants: a Standard model optimized for visual quality and creative control, and a Fast version intended for rapid iteration and high-volume production. It features enhanced prompt adherence, allowing for precise control over camera movement and scene composition.