Paper summary: Qwen-Image Technical Report

Problem

Image generators often garble any text they draw, especially long text and logographic scripts like Chinese. Image editing also struggles to change one thing while keeping the rest consistent.

Method

• Text rendering: a full data pipeline (collection, filtering, annotation, synthesis, balancing) plus a curriculum that starts with no-text images, moves from simple to complex text, and scales to paragraph-level descriptions.
• Editing: multi-task training on text-to-image, text-image-to-image and image-to-image reconstruction, which aligns the latent spaces of Qwen2.5-VL and the MMDiT generator.
• Dual encoding: the original image is fed to Qwen2.5-VL (semantic representation) and to a VAE encoder (reconstructive representation), so edits keep the meaning and the pixel detail.

Key results

From the abstract: strong performance on alphabetic languages like English, “remarkable progress” on logographic languages like Chinese, and state-of-the-art results in both generation and editing across multiple benchmarks.

Limitations

There’s no dedicated limitations section in the extracted text. Benchmark comparisons are the developers’ own.