Wan 2.7 Image for Editing and Sets

Overview

Wan 2.7 Image by Alibaba is a single model for image generation and prompt-driven image editing. You write one instruction, and it produces one or more finished PNG images. You can also upload up to 9 reference images and tell the model what to keep, swap, or blend. This helps you get consistent creative results without hopping between separate generation and edit tools.

What you can build

• Text-to-image concept art for ads, posters, and social creatives
• Prompt-based photo edits like weather changes, object swaps, and style changes
• Multi-image fusion like “use the product from image 1 in the scene of image 2”
• Coherent image sets for storyboards, tutorials, or “same character across scenes”
• Consistent variations for product shots, angles, and backgrounds

Inputs

• A required text instruction in English or Chinese, up to 5,000 characters. For multi-image workflows, reference your uploads as “image 1”, “image 2”, and so on.
• Optional reference images to edit, restyle, or blend. You can provide 0 to 9 images.
• Reference image file constraints: JPEG/JPG, PNG (no alpha channel), BMP, or WEBP. Each image must be 240 to 8,000 pixels on each side, under 20 MB, with aspect ratio between 1:8 and 8:1.
• An output size choice for the generated images. Typical tiers are 1K and 2K, and some deployments also accept custom width×height sizes.
• A requested number of output images. Standard generation returns up to 4 images per request. Coherent image-set generation can return up to 12 images.
• An optional “coherent set” toggle that makes the model produce a single consistent series instead of separate takes.
• An optional “thinking” toggle that can improve quality for pure text-to-image. It only applies when you do not provide reference images and you are not generating a coherent set.
• An optional random seed value for repeatable outputs. Valid range is 0 to 2,147,483,647.

Outputs

• One or more PNG images.
• When you provide reference images, the output keeps the same aspect ratio as the last provided reference image and scales to the selected size tier.
• When you provide no reference images, the output is typically a square image at the selected size tier.
• If you request multiple outputs, you receive a list of PNG images, one file per generated result.

Recommended settings

• For the best text-to-image quality: don’t upload reference images, keep coherent-set mode off, and leave thinking enabled.
• For a consistent storyboard or tutorial sequence: enable coherent-set mode and describe each frame in order (for example, “First image: … Second image: …”).
• For controlled edits: upload 1 to 9 reference images and give one direct edit instruction that names which image supplies which element.
• For reproducible runs: set a seed and keep the same prompt, size tier, and reference images.

Limitations

• Wan 2.7 Image outputs up to 2K resolution. 4K output is a Pro-only feature and applies only to text-to-image.
• Coherent image sets and image editing are also capped at 2K output.
• The model supports only a single user turn per request. It is not a multi-turn chat loop.
• Prompts longer than 5,000 characters are truncated.
• PNG reference images cannot include transparency. Alpha channels are not supported.
• Low-quality inputs can reduce edit quality. Blurry images, heavy compression, extreme crops, or inconsistent lighting make blends look less natural.

Safety & compliance

• Follow Alibaba Cloud Model Studio terms and all applicable laws for your region.
• Don’t try to bypass safety controls or platform restrictions.
• Don’t use the service to fabricate or spread false information, or other prohibited content.
• Some publishing contexts require clear labeling or marking of AI-generated outputs. Add watermarks or disclosures when required.
• Alibaba’s terms also restrict using model access or outputs to train or develop competing services without authorization.