Qwen Image 3.0 Image Gen and Editing

Overview

Qwen Image 3.0 is Alibaba's image model for generation and instruction-based editing. You provide a long, detailed prompt and it produces a PNG image with dense layouts and readable on-image text. If you upload 1 to 3 reference images, it uses them as visual anchors while applying your edit instructions. This is useful when you need posters, UI screens, or storyboards in one pass.

What you can build

• Posters, flyers, and infographics with legible text and structured layout
• Multi-panel storyboards and shot lists rendered as a single image
• UI mockups that resemble real app or web screens
• Exam papers, worksheets, and slide-like visuals with formulas and diagrams
• Style changes and scene re-lighting for an existing image
• Object insertion, removal, or replacement described in natural language
• Reference-guided concepts where one image sets subject and another sets style

Inputs

• A single instruction prompt that describes the image you want, including layout and any on-image text. The model is designed to accept prompts up to about 4,500 tokens.
• Optional reference images you upload (up to 3). Use them for edits, restyling, or multi-reference fusion. Common formats include JPG/JPEG, PNG, WebP, BMP, TIFF, and GIF. Keep each file at 10 MB or less. For best results, keep image width and height within 384 to 2048 pixels.
• An output quality tier selection (1K or 2K). Use 2K when you need smaller, cleaner text or more fine detail.
• An aspect ratio selection (or automatic). Auto follows the first reference image when you provide one. If you provide no reference images, auto defaults to a square.
• The number of output variants to generate in one run (1 to 6).
• An optional negative instruction that lists elements you want to exclude from the result. If you use one, keep it short. Some providers cap this at about 500 characters.
• Optional prompt rewriting controls that can expand short prompts into more specific instructions. Some modes only work for text-to-image, not when you include reference images.
• An optional numeric seed to make runs more repeatable (0 to 2,147,483,647).

Outputs

• One to six generated images.
• Each result is a PNG image.
• Output resolution stays within a total pixel range equivalent to 512×512 up to 2048×2048.
• Output aspect ratio stays within roughly 1:8 to 8:1.
• When you use reference images, the outputs are edits or transformations that follow your instruction while keeping key visual cues from the references.

Recommended settings

• Turn on prompt rewriting when your prompt is short or underspecified. It can improve detail and layout fidelity.
• Use the "direct" rewrite style for most prompts. Use the heavier "agent" rewrite style only for pure text-to-image runs.
• Enable thinking mode only when prompt rewriting is enabled. It is designed to improve output quality.
• Pick the 2K tier for typography-heavy work like posters, menus, UI, and documents.
• Keep aspect ratio on auto for image edits when you want to preserve the reference image framing.

Limitations

• The model supports up to 3 reference images per request for guided editing or fusion.
• Output is limited to PNG and up to 2048×2048-equivalent pixel counts.
• Very wide or very tall compositions are limited. Keep aspect ratios within about 1:8 to 8:1.
• Edits are instruction-based. This workflow does not rely on a separate mask input.
• Some rewrite modes are not available when you provide reference images.
• Prompts beyond the supported limit may be truncated, which can drop important constraints.
• Low-quality inputs can reduce edit accuracy. Blurry photos, heavy compression, watermarked images, and upscaled images often cause drift.
• Dense on-image text can still contain spelling mistakes, missing glyphs, or uneven kerning. Proofread before publishing.

Safety & compliance

• Alibaba's Qwen policy forbids harmful uses, including sexual content involving minors, privacy violations, violence or terrorism promotion, hate or discrimination, fraud, and attempts to bypass safeguards.
• Avoid uploading personal data you do not have rights to share. Do not upload minors' personal data without legal guardian consent.
• Alibaba Cloud Model Studio states it does not use your data for model training. It also states data transmission is encrypted with AES-256.
• You are responsible for having the rights to any reference images you upload and for reviewing outputs before use in high-impact contexts.