Qwen Image guide: how to create accurate text, edit images, and combine references

Qwen Image is an image generation and editing model from Alibaba’s Qwen team. Its original open-weight release combines visual generation, text rendering, and instruction-based editing in one model, with particular strength in Chinese and English typography.

In Phygital+, the Qwen Image node can start from a text prompt or use up to three images as references. Those references can define the subject, product, style, layout, or environment. The node also exposes LoRA support, multiple aspect ratios, seed and guidance controls, and camera presets.

This guide focuses on the Qwen Image implementation available in Phygital+. Newer Qwen Image 2.0 and 3.0 families now exist, so version-specific functions are separated instead of being presented as one feature set.

Qwen Image versions compared

What Qwen Image is good at

Text-to-image generation

The model handles detailed descriptions of subject, composition, style, lighting, and typography. It is useful for posters, illustrations, product concepts, editorial images, and structured visual assets.

Instruction-based editing

Connect an existing image and describe what should change. The clearest edit prompts specify the target region, the new content, and the elements that must remain untouched.

Multi-image composition

Up to three source images can contribute different parts of a result. One can define a person, another a product, and a third a location or style. The prompt should assign a clear job to every image.

Multilingual text

The original model was designed around complex text rendering, especially Chinese and English. Other scripts can work, but current official editing documentation guarantees Simplified Chinese and English, so every localized result requires proofreading.

Camera and LoRA control

The Phygital+ node exposes camera angle and distance controls as well as trained and preset LoRAs. These controls make it possible to explore composition and visual style without rewriting the entire prompt.

How to work with multiple references

Reference order has no universal meaning by itself. Define each role in the prompt:

• Image 1: preserve the person’s face and hairstyle.
• Image 2: use the jacket and accessories.
• Image 3: place the subject in this interior and match its lighting.

When an edit fails, remove references that describe the same element differently. A concise set of compatible inputs is easier for the model to reconcile.

Generation versus editing

Without a start image, the node generates from the prompt. With references, the task becomes editing or composition. In edit mode, describe the transformation and preservation rules instead of repeating every visible detail.

How to generate text inside images

1. Put exact copy in quotation marks.
2. Identify the language separately from the prompt language.
3. Describe text position, size, and hierarchy.
4. Limit the first generation to the essential headline and one supporting line.
5. Request no additional words.
6. Proofread the final result at full size.

Qwen Image can build posters and layouts around longer copy than many diffusion models, but small print, mixed scripts, unusual names, and dense paragraphs still require manual review.

How to prompt Qwen Image

Write the prompt as a production brief: output type, subject, composition, exact text, style, lighting, reference roles, and preservation constraints.

Qwen Image prompt examples

Building a Qwen Image workflow in Phygital+

Qwen Image can sit in the middle of a broader design pipeline. A text model can structure the brief and exact copy. Image nodes can create the subject, product, or environment references. Qwen combines or edits them, while branches test camera, layout, LoRA, and guidance settings. The approved output can continue into upscaling or video.

1. Define the deliverable, copy, and preservation rules.
2. Create or collect up to three references with distinct roles.
3. Generate or edit in Qwen Image.
4. Branch camera angles, compositions, or LoRA styles one variable at a time.
5. Check text, identity, product details, and shadows.
6. Upscale or animate the selected image.

The canvas preserves the relationship between every source and variation, which makes client revisions easier to reproduce.

Qwen Image pricing in Phygital+

Qwen Image uses Phygital credits. The current cost appears inside the node before generation and can depend on the active options and image count. The same subscription balance can be used for the surrounding text, image, video, and utility nodes.

Common Qwen Image problems

The edit changes too much

Name the exact target and list every important element that must remain unchanged.

References conflict

Assign one role to each image and remove duplicates that disagree about identity, product shape, or style.

Text is misspelled

Shorten the copy, use quotation marks, request no extra words, and proofread at full resolution.

The composite looks pasted together

Ask the model to match perspective, light direction, shadows, color temperature, and depth of field.

The character loses identity

Use the clearest face as the primary reference and avoid simultaneous changes to face, hair, pose, camera, and style.

Camera controls distort the subject

Make smaller angle changes and keep the subject or product constraints explicit.

Current limitations

Dense text, mixed-language layouts, tiny lettering, exact logos, complex hands, and aggressive viewpoint changes can still fail. LoRAs and references can also pull the image in competing directions. Treat generated trademarks, packaging copy, and regulated claims as drafts until they are reviewed.