Qwen: Qwen Image 3

Qwen-Image-3.0 is Qwen’s third-generation image model, combining text-to-image creation and reference-based editing in one model. The standard edition balances image quality and response speed, while Qwen-Image-3.0 Pro is the flagship quality-oriented sibling. A request produces an image: a poster, illustrated explanation, product scene, storyboard or revised photograph rather than an editable document or a running interface.

Its rich-content capabilities suit a canvas with several related sections or a nested visual layout. Prepare the headings, exact wording, hierarchy and relationship between panels before adding stylistic detail. For a product comparison, a useful brief specifies the main product view, the attributes to label and the reading order of the supporting panels. For a storyboard, describe each panel’s subject, action and caption separately. The recommended prompt maximum is 4,500 tokens; spend that space on the content and layout the image actually needs.

The model supports detailed textures, typography down to small text demonstrated at 10 pixels, and native rendering across 12 languages with multiple fonts and artistic styles. Those capabilities can combine a realistic object with an information-rich graphic rather than forcing a choice between a photograph and a text-heavy layout. Supply the exact words and numerals to render, their intended language and where they belong. Treat a generated interface as a visual mockup: buttons, charts and labels in the image do not implement the behavior of an application.

Editing can change text, add or remove objects, move elements, adjust a subject’s action, transfer a style or enhance a damaged image’s details. Supply one to three images and a single clear editing instruction. Image numbering follows the supplied order, so identify which image contributes the subject, which provides a garment or object and which supplies the setting. For a product scene, ask to place the object from image 1 into the environment from image 2 while preserving the features that distinguish the product. Keep changes and preservation requirements explicit.

The exact image-input workflow accepts JPG, JPEG, PNG, BMP, TIFF, WebP and GIF; only the first frame of an animated GIF is processed. Each image can be up to 10 MB. The current API accepts dimensions from 1 to 8,000 pixels with no more than a 10:1 long-to-short-side ratio, and recommends 384–2,048 pixels for good results. Choose an input that clearly shows the details being transferred or edited. A tightly cropped reference can obscure the surrounding structure needed to place an object naturally.

The output is PNG, with one to six images per request and one as the default. Output pixel area ranges from 512 × 512 to 2,048 × 2,048, with aspect ratios between 1:8 and 8:1; these are area limits, not a rule that every side must be at most 2,048 pixels. If no size is supplied, the model recommends a resolution from the prompt. Select a landscape, square or portrait composition for the intended use, then inspect the returned image at that actual size so tiny labels and fine detail can be assessed where they will appear.

Prompt rewriting is enabled by default in the producer workflow. Direct enhancement works for both generation and editing; agent enhancement is a more elaborate rewriting mode for text-to-image only and is rejected for image-to-image editing. Thinking is enabled by default, requires prompt rewriting and increases generation time. Negative prompts can describe unwanted elements, while a fixed seed helps compare variations. State exclusions and exact preservation requirements clearly rather than expecting a seed to make a changing brief return an identical picture.

For a useful final artifact, inspect the rendered words, units, panel order and details that identify the subject, then revise the specific part that missed the brief. A multi-image response gives alternatives to compare; keep each returned image associated with the prompt and references that produced it. The producer’s image locations expire after 24 hours, so retain the actual image asset when the result becomes part of a project. The standard 3.0 edition remains distinct from Pro and from Qwen 2.1 models with their different reference counts and transparency features.