Image · alibaba/qwen-image-3
Qwen Image 3 — images with readable text in the frame
Qwen Image 3 is the next generation of Alibaba's image model. The Qwen-Image line is known for rendering in-frame text — signs, packaging, posters — more accurately than most competitors, and the third version continues that specialty. On limko the model works both ways: text-to-image and editing an existing picture from a reference.
from 5 ₽ per image
Where Qwen Image 3 excels
- Frames with lettering: a shop sign, a label, a poster — where other models smear the letters.
- Reference-based editing: upload a picture, describe the change — match_input_image keeps the aspect of the original.
- Non-standard formats: nine aspect ratios including wide 2:1 and vertical 1:2 for banners and stories.
Qwen Image 3 or Nano Banana / GPT Image
Google's Nano Banana is the everyday workhorse for photo edits and character series. Pick Qwen Image 3 when the lettering in the frame matters: typography is the line's specialty, and it holds up better on signs, packaging and posters.
Nano Banana Pro is also strong with text and the most reliable with Cyrillic, but noticeably pricier per run. At 5 ₽ per frame, Qwen Image 3 is an inexpensive way to get a frame with readable letters; for a final cover with dense layout, still compare both models on your prompt.
What you need as input
A text prompt; for editing — a reference image. Negative prompt (what to keep out of the frame), seed for reproducible series and prompt expansion are available — the model fleshes out a short description on its own.
Aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1 and 1:2, or match_input_image to repeat the reference format. One run — one image, the price is shown on the button before you click.
Where the model is limited
One run returns one frame — for a series of variants, launch several generations with different seeds. Long phrases with complex layout are safer as short wording: that is the general rule for in-frame text with any model.
Qwen Image 3 FAQ
How does the third version differ from the previous Qwen-Image?
It is the next generation of the line: closer prompt following and more stable reference-based editing. The previous version stays in the catalog — a series started on it can continue without a style shift.
Does it understand non-English prompts?
Yes, it handles them, and it can render non-Latin lettering in the frame. For fine style wording English remains the safer choice — the general rule for catalog models.
How does reference-based editing work?
Upload the source image and describe the change. The match_input_image mode keeps the source aspect, and the negative prompt lets you explicitly forbid the unwanted.
How much does a generation cost?
About one and a half credits per frame — the exact price is always visible in the interface before launch.