Alibaba has just launched Wan 2.7 Image, a unified generation and editing model that solves the biggest problem of artificial intelligence: identical faces, messy colors, and broken text.
Most AI image generators have a "tell" function. Faces appear overly smooth, colors deviate from your request, and any text exceeding two words turns into abstract art. Wan2.7-image addresses all three issues simultaneously—at least based on our tests so far, the results are truly impressive.
It doesn't look like the same person's face
The biggest visual feature of portrait paintings generated by artificial intelligence is similarity. By exchanging hair and clothes, you will swear that every face comes from the same mold. Wan2.7-Image directly introduces granular facial carving in the prompt: you can specify the skeletal structure, facial shape (circular, square, rectangular), eye style (narrow, deep, wide), and other subtle features.
The results are not only 'different', but also credible. Faces appear to belong to actual, specific individuals, rather than a mixed average. For anyone creating character art, storyboards, or brand personas, this is an important step towards output that can be used without manual retouching.
Point and edit
Wan 2.7 Image Palette: What Precision Designers Really Need
If you have ever tried to match the exact color scheme of a brand in an AI generated image, you will know the pain involved: the model will interpret 'navy blue' in any sense, and the output is rarely consistent with your brand guidelines.
Wan2.7-Image comes with a palette extraction feature. Put in a reference image - a mood board, a painting, a screenshot of a design system - and the model will extract the color distribution and apply it to the generated output. You keep your essay and theme; The color palette will move to match the reference.
For designers and marketing teams working under strict brand guidance, this eliminates a whole round of post production color correction. We tested it with Pantone color cards and Van Gogh's paintings - both of these transfer prints are very faithful.
Wan 2.7 Image Text Rendering: 3K Token, Full Page
Text rendering has always been a fatal weakness in image generation. Perhaps most models can handle a short title. Ask a paragraph and you will get a beautiful image filled with beautiful yet meaningless waves.
Wan2.7-Image pushes the boundary to 3000 tags, enough to fill the entire A4 page. We are discussing generating images of research papers, data tables, mathematical formulas, and dense infographics, where each character is clear, easy to read, and positioned correctly. This model supports 12 languages, including English, Chinese, Japanese, Korean, and major European languages.
We threw it a whole page of product specification sheet - finely printed, project list, footnotes - and the output is ready for printing. This is the first time for us.
Point and edit
Generations are only half of the story. Wan 2.7 Image natively supports interactive editing: Select an area where you can add, replace, move, or realign elements without leaving the model or exporting to a separate tool.
This is just like the feeling you expect from a layer based editor, except there are no layers - only natural language instructions applied to bounding boxes. Move the coffee cup to the left. Replace the background with an outdoor caf é. This quickly blends seamlessly with the rest of the image.
Up to 9 references with multi topic consistency
Generating a great image is one thing. Maintaining consistency in characters, products, or styles across a series of images is where most models collapse.
Wan 2.7 Image supports up to 9 reference images to achieve theme consistency. Feed your hero character from multiple angles, and it will maintain its identity in new scenes, poses, and compositions. This is of great significance for the following aspects:
E-commerce product photography (same model, different clothing/settings)
Storyboard and Manga Generation (same characters, different panels)
Brand activities (consistent visual identity across assets)
