Qwen Image

Qwen Image is a 20 billion parameter image foundation model from Alibaba's Qwen team, released in August 2025 under the permissive Apache 2.0 license. It uses a Multimodal Diffusion Transformer (MMDiT) architecture: the Qwen2.5-VL vision-language model reads and understands your prompt, and the diffusion transformer turns that understanding into pixels. A positional encoding scheme called MSRoPE lines up text and image information inside the model, which is part of why it scales well to different resolutions.

Its signature skill is rendering readable text inside images, from single words on a sign to multi-line paragraphs on a poster, in both English and Chinese. The team trained this deliberately with a curriculum approach, starting with images containing no text, then simple captions, then gradually scaling up to paragraph-level text, backed by a large pipeline for collecting, filtering, annotating, and synthesizing training data. On release it topped public text-to-image benchmarks including GenEval, DPG, and OneIG-Bench.

The same foundation powers editing, not just generation. A multi-task training setup lets companion Qwen-Image-Edit models change objects, transfer styles, or rewrite text in an existing picture while keeping the rest of the image intact. Because the weights are open, the model also runs locally through ComfyUI, and quantized community versions bring it within reach of consumer GPUs with roughly 8 to 16 GB of VRAM.

The model line has kept moving. A December 2025 update (Qwen-Image-2512) focused on more natural-looking people, with visible skin pores and shading instead of the smoothed, artificial look users had flagged. In February 2026 the team released Qwen-Image-2.0, a leaner 7 billion parameter model that generates at native 2K resolution, accepts prompts up to about 1,000 tokens for text-heavy designs like infographics and slides, and merges generation and editing into a single model.