AlibabaのQwen-Image-2.0、圧縮率を2倍に
AlibabaのQwen-Image-2.0は、VAEディスクリミネータを廃止し新しいプロンプト拡張モジュールを採用することで、16倍の空間ダウンサンプリングを達成し、生成ステップを4に削減した。
alibaba's technical report on qwen-image-2.0 outlines architectural improvements for image generation, specifically a vae that utilizes 16-fold spatial downsampling compared to the standard 8-fold ratio found in models like flux.1-dev. this compression increase accelerates training and inference but typically risks detail loss; the qwen team mitigated this using skip connections and shaped latent spaces without a discriminator. the architecture also replaces standard feed-forward blocks with swiglu to prevent 'massive activations' during joint text-image training. the system incorporates a prompt expansion module powered by qwen3.5-9b trained on reverse-engineered data to convert short user inputs into rich descriptions.