Wan 2.7-Image Complete Guide: Face Customization, Color Palette, Text Rendering & More

Wan 2.7-Image dropped on April 1, 2026, and it's a different beast from previous Wan models. Instead of just generating pretty pictures, this is a unified generation-and-editing model. Here's how to actually use each feature.

Face Customization — The 'Thousand Faces' System

The old problem with AI portraits: everyone looks the same. Wan 2.7-Image fixes this with granular facial controls. In your prompt, you can now specify bone structure (high cheekbones, strong jawline), eye shape (almond, deep-set, phoenix, round), and face shape (oval, round, square, rectangular). The model doesn't just swap templates — it adjusts the underlying geometry. Tip: combine multiple attributes for unique results. 'Round face with deep-set eyes and high cheekbones' produces something genuinely distinctive.

Color Palette Control

Upload any reference image and Wan 2.7-Image extracts its exact color distribution — not just the dominant color, but the full ratio of every hue. You can then apply this palette to new generations. Use cases: reproducing a client's brand colors exactly, matching a movie's color grade, or keeping a series of illustrations visually consistent. The model preserves your composition while swapping the entire color space.

3K Token Text Rendering

Previous models mangled text after ~50 characters. Wan 2.7-Image handles up to 3,000 tokens across 12 languages with print-quality clarity. That's enough for a full A4 page of academic text, complex mathematical formulas, or data-heavy tables. The secret is the unified architecture: text and image share the same latent space, so the model 'understands' what it's writing rather than treating text as visual patterns.

Interactive Editing

Don't like something in your generated image? Draw a box around it and type what you want instead. Add elements, move objects, align compositions — all with pixel-level precision. This isn't inpainting; it's semantic editing. The model understands spatial relationships, so moving an object also adjusts shadows and reflections.

Multi-Subject Consistency (Up to 9 References)

Feed the model up to 9 reference images, and it maintains character identity and style across all outputs. Essential for storyboards (same character across scenes), e-commerce (same model in different outfits), and architectural visualization (same building from different angles).

Batch Generation — Up to 12 Images

Generate up to 12 images in a single request. All images share the same style and quality level, making this perfect for PPT slide illustrations, social media content calendars, or product photography series.

Pro Tip: Wan 2.7-Image-pro

If you need even more stable compositions and better semantic understanding, try Wan 2.7-Image-pro. It's trained on a larger dataset and available through the same API endpoint. Just switch the model parameter.

Getting Started

Try it at tongyi.aliyun.com/wan or integrate via the Alibaba Cloud API. Wan 2.1 remains fully open-source if you prefer local execution with no content restrictions.