Qwen-Image-2.1 Prompt Guide

Qwen-Image-2.1 (released 2026-09-20) is a 7B unified model: text-to-image generation, image editing, multi-reference compositing, transparent (RGBA) output, and best-in-class text rendering, at native 2K resolution. It is prompted with fluent natural language, not Stable Diffusion-style tag lists.

Core rules (always apply)

1. Natural sentences, never tag lists. Qwen image models ignore SD-style tags and (term:1.5) weight syntax. Emphasize by description instead: "with vibrant, striking red hair".
2. Write descriptive prompts in English, whatever language the user speaks. Only text that must appear inside the image keeps its original script (Japanese stays Japanese, Chinese stays Chinese).
3. Front-load the subject. The transformer weights early, concrete tokens heavily — subject first, then environment, style, composition, lighting.
4. Any string that must appear in the image goes in double quotes, verbatim. a neon sign that reads "GRAND OPENING". Case and punctuation are reproduced as written.
5. Edit prompts state the edit + a preserve clause. "Change X to Y. Keep everything else unchanged." One logical edit per pass; chain small passes for big changes.
6. Pass aspect ratio / resolution as a parameter (wh_ratio, size, or your UI's ratio picker) — never write "16:9" or "4K" inside the prompt text.
7. No quality boosters. "masterpiece", "8K", "highly detailed", "award-winning" are officially discouraged for 2.1 — describe what is visible instead.

Routing

The user wants…	Read before writing
A fresh image from a text description	references/text-to-image.md
To modify, restyle, or combine existing images	references/image-editing.md
Readable text inside the image (posters, signs, UI, charts)	references/text-rendering.md (plus the T2I or editing page)
Transparent PNG, sticker, or subject cutout	references/capabilities.md → Transparent images
Aspect ratios, negative prompts, CFG, seed, steps, or how to run the model	references/capabilities.md
Concrete prompts that are known to work	references/examples.md

Workflow

1. Identify the mode: generate, edit, or composite. If readable in-image text is involved, the text-rendering rules apply on top of the mode's rules.
2. Read the matching reference file. Do not write from memory — the official style rules are detailed and counterintuitive in places (e.g. long observational paragraphs for layouts, preserve-clauses for edits).
3. Draft the prompt in English. Check it against the core rules above.
4. Choose the aspect ratio semantically (portrait subject → 2:3, presentation → 16:9, badge/icon → 1:1) and pass it as a parameter. Default 3:2 landscape / 2:3 portrait.
5. If the user reports a failure (garbled text, wrong subject, drifted face), consult the failure-pattern table in references/capabilities.md before rewriting.

Minimal templates

Simple generation:
A neon shop sign that reads "QWEN IMAGE 2.1", rainy night, reflections on wet pavement

Simple edit:
Change the background to a sunset beach. Keep the subject, pose, and lighting unchanged.

Transparent sticker:
This is an RGBA image with transparency. A cute cartoon dragon sticker.
The image has alpha channel and the background is transparent.

For anything with multiple elements, precise placement, or several text blocks, escalate to the long-form style in references/text-to-image.md — short prompts underperform on layouts and posters.