ಕ್ವೆನ್ ಇಮೇಜ್ 2.0 ಪ್ರೊ

Alibaba's flagship Qwen image model, the highest-ranked of the models we route through Together AI. It is notably strong at rendering text inside images, in both Chinese and English.

Qwen had been developing two separate lines of image model: a generation track that chased text rendering and then photorealism, and an editing track that went from single-image edits to multi-image edits to consistency between them. This release merges the two into one model that does both jobs in the same mode. That is the reason the editing side improved as much as it did - every gain in text rendering and realism on the generation side carries straight into an edit, instead of needing to be reproduced in a separate model.

Typography is the headline claim. The model accepts instructions of around a thousand tokens and renders professional infographics, slides, posters and comics from them, and the team's demonstration was to generate the launch post's own timeline slide with the model, complete with picture-in-picture panels that stay visually consistent with one another. Text stays aligned inside calendar grids, table cells and comic speech bubbles, and when text and imagery share a frame the model tends to place the words in empty areas rather than over the subject.

The realism claim is about where text sits, not just whether it is spelled right. Qwen shows the model writing on a glass whiteboard, on a printed shirt and on a magazine cover in the same scene, with each surface's lighting, reflection and perspective respected. It handles both Chinese and English, and several distinct calligraphic styles; the team's own honest note on a long classical passage in small script is that nearly all of it comes out correct with a handful of characters imperfect, which is a fair statement of where very dense small type starts to break down.

Away from text, the lab reports native 2K output with fine detail across skin, hair, fabric, architecture and foliage - its example claims more than 23 distinguishable shades of green in a single natural scene - and describes the architecture as lighter than its predecessors, smaller and faster to run than the two-track lineage it replaces.

Here the model takes up to three reference images, as PNG, JPEG or WebP up to 20 MB, and edits are prompt-driven: there is no mask, so which region changes is decided by the model from your wording. Sizes run from 1024 by 1024 up to 2048 by 2048 at a flat per-image price, and there is no seed, so the same prompt will not reproduce the same image. It earns its rate on pictures that contain words - a poster, a slide, a label, a chart - and a plain photographic prompt does not need what it is best at.