analiza aiqbench

Alibaba's release of Qwen-Image-2.1 consolidates text-to-image generation and editing into a single 7B diffusion transformer, significantly reducing the computational overhead previously required by separate checkpoints. The architecture leverages mixed-granularity attention to cache prefix keys and values, enabling efficient processing of multi-reference edits without re-computing condition tokens at every denoising step. This unified approach offers a cost-effective alternative for deploying image generation workflows that require native RGBA transparency and local editing capabilities.