Wan-2.2 Image

The Pro (14B) variant uses a 27B Mixture-of-Experts architecture (14B active parameters per step) to deliver maximum prompt compliance, complex multi-character staging, and intricate textures like fabric weaves and fine hair. The Lite (5B) variant is optimized for rapid, cost-effective iteration and drafting, producing coherent layouts at a lower compute cost with slightly less high-frequency surface detail.

Use the Lite (5B) model for rapid storyboarding, layout exploration, and initial prompt testing where generation speed and draft composition matter most. Switch to the Pro (14B) model for final production renders that require exacting anatomical accuracy, complex multi-subject interactions, and subtle photographic lighting falloff.

The model excels at photorealistic portraiture, cinematic film stills, complex environmental scenes, and anatomically precise hands. Its video-derived architecture enforces natural physical logic, realistic skin blemishes, and nuanced lens depth of field while bypassing the synthetic, oversaturated CGI look common to other generators.

Avoid using Wan-2.2 Image for flat vector art, typography, graphic logos, or minimalist line drawings. The model is natively optimized for photographic realism and detailed compositional depth, and it can produce sterile results when prompted with ultra-brief, unguided text.

Guidance scale is best kept between 3.0 and 5.0 (default 3.5) for balanced natural contrast without color oversaturation. The shift parameter controls denoising dynamics: lower values (2 to 4) introduce dynamic lighting and dramatic surface textures, while higher values (7 to 10) produce smoother, more predictable finishes.

Wan-2.2 Image was developed by Alibaba Tongyi Lab and released in July 2025, adapting the open-source Wan-2.2 video foundation model family into a dedicated high-resolution static image generator.