Alibaba Qwen Image 3.0 Pro

Overview

Qwen Image 3.0 Pro is an image generation and image editing model by Alibaba (Qwen). It turns a long natural-language instruction into a high-resolution PNG image, or edits an existing image using 1–3 references plus a single instruction. It’s designed for information-dense images with readable text, like posters, menus, and UI mockups. It helps when you need images that carry real content, not just style.

What you can build

• Posters and flyers with small, legible text
• Menus, price lists, and product one-pagers with structured layouts
• Multi-panel storyboards and comic-style pages
• UI mockups and “screenshots” for web, game, and livestream scenes
• Product images that include packaging text and labels
• Identity-preserving photo edits, like outfit changes or scene swaps
• Style transfers and restyling using a reference image
• A/B variants for creative testing (up to 6 outputs per run)

Inputs

• A required instruction prompt, written as normal sentences. The service recommends staying within about 4,500 tokens.
• Optional reference images for editing or reference-guided generation. You can provide 1 to 3 images.
• Reference images can be provided as a public HTTP(S) URL or as Base64 in a Data URL format.
• Supported reference formats include JPG, JPEG, PNG, BMP, TIFF, WEBP, and GIF.
• Each reference image can be up to 10 MB. For best results, keep width and height between 384 and 2048 pixels.
• An output size you choose, either automatic sizing or an explicit width × height. Total pixels must stay between 512×512 and 2048×2048. Aspect ratio must be between 1:8 and 8:1.
• A choice of output quality tier (for example, a 1K tier vs a 2K tier) when the UI offers it.
• How many images to generate in one request (1 to 6).
• An optional negative instruction that lists elements you want excluded.
• An optional prompt-rewrite toggle that expands and refines your instruction before generation.
• An optional prompt-rewrite style. A “direct” rewrite works for both generation and editing. A stronger “agent” rewrite only works for text-to-image.
• An optional “thinking mode” toggle for higher quality. It only works when prompt rewriting is enabled.
• An optional random seed (0 to 2,147,483,647) to make results more repeatable.
• An optional watermark toggle.

Outputs

• One to six generated images in PNG format.
• Results are typically delivered as hosted image URLs that expire after about 24 hours, so save outputs you want to keep.
• Output metadata may include the final width and height, plus the input and output image counts.

Recommended settings

• Keep prompt rewriting enabled when your prompt is short or underspecified. This is the default.
• Use the “direct” rewrite style for most work, including edits with reference images.
• Use the “agent” rewrite style only for pure text-to-image runs. Don’t use it when you upload reference images.
• Leave thinking mode enabled for best quality, but only when prompt rewriting is on.
• Turn prompt rewriting off when you need strict control over exact wording and composition.
• Set a seed when you want to re-run a prompt and compare changes.

Limitations

• Output images are limited to PNG.
• Maximum output resolution is capped by the 2048×2048 total-pixel constraint.
• Editing accepts 1 to 3 reference images per request.
• Mask-based edits (inpainting with an explicit mask image) aren’t supported in the Qwen Image 3.0 API flow.
• Agent-style prompt rewriting isn’t supported for image-to-image editing.
• Some OpenAI-compatible integrations ignore Base64 response requests and still return image URLs.
• If you upload an animated GIF, the service processes only the first frame.
• Low-quality reference inputs can reduce edit fidelity. Blurry faces, heavy compression, or tiny subjects can break identity preservation.

Safety & compliance

• Follow Alibaba Qwen usage rules. Don’t request illegal, exploitative, hateful, or violent content.
• Only upload reference images you have rights to use. This includes people’s likeness and copyrighted material.
• Don’t submit sensitive personal data, confidential documents, or regulated data in prompts or reference images.
• If you enable watermarking, don’t remove or alter any watermark applied to outputs without permission.