Alibaba's Wan2.7-Image AI Generates Custom Faces and Writes Essays

Today, Alibaba officially launched its unified large model for image generation and editing, Wan2.7-Image. This model not only achieves a qualitative leap in visual quality but also overcomes traditional AI image generation limitations, such as "generic faces" and "prompt misalignment," through comprehensive capability upgrades.

Farewell to AI-Generated Faces: Entering the Era of "One Person, One Face"

Wan2.7-Image significantly enhances its virtual character customization function. Users can customize everything from bone structure and eyes to subtle facial features, precisely controlling specific traits like an oval face, phoenix eyes, or deep-set eye sockets. This advancement moves beyond the mechanical uniformity of past AI portraits, enabling true personal expression.

"Color Palette" Feature and "Print-Quality" Text Rendering

In artistic expression, the model now supports a "Color Palette" feature. This allows users to extract the color composition from a reference image—such as Matisse's red series or Van Gogh's yellow series—with a single click and accurately apply it to new creations. Furthermore, Wan2.7-Image excels at long-text rendering, supporting inputs of up to 3K tokens. It can stably output an entire A4 page of content containing complex formulas and tables, meeting print-quality standards across 12 supported languages.

Interactive Editing and Multi-Subject Consistency

The model boasts powerful interactive editing capabilities, supporting the alignment, movement, or replacement of elements via precise selection. For instance, users can select characters in an image to swap their positions or replace ice cubes with fruit, achieving pixel-level control. Simultaneously, the model supports multi-subject consistency across up to 9 images, maintaining a uniform style and characteristics when generating AI girl groups or furniture sets.

Core Technological Breakthroughs and Industry Applications

Wan2.7-Image employs a leading unified architecture for both generation and understanding, achieving semantic mapping within a shared latent space. This means the model no longer merely guesses text to match pixels but possesses a foundational semantic understanding. The model has launched alongside the Wan2.7-Image-pro version, which offers more stable composition and precise comprehension.

This model is now widely applied in short-form video production (one actor playing multiple roles), e-commerce advertising (one model image for multiple uses), education, research, and social entertainment. Users can access the API via the Alibaba Cloud BaiLian Platform or experience it directly on the Wanxiang Official Website.