Qwen-Image-2.1 review: AI images with real text rendering

Alibaba released Qwen-Image-2.1, a ~20B-parameter image model with improved text rendering in Chinese, English, Japanese and Korean. It supports reference-based editing, outpainting and 1080p output, generating a 1024x1024 image in about 10 seconds on a modern GPU.

• ~20B-parameter model with a text encoder and word localization
• Short phrases render reliably, long paragraphs turn into gibberish
• Reference-based editing changes the target element but often alters lighting and background
• Full model needs heavy VRAM; quantized version runs on a 24GB card