Alibaba Qwen-Image-2.0 uses a 16× compression VAE and four generation steps
Qwen-Image-2.0 is a multimodal diffusion transformer with Qwen3-VL as its conditioning encoder. Its VAE downsamples spatially by 16×, and distillation cuts generation from 40 steps to 4. It reached an Elo score of 1,168 on LMArena. The timing of an official weight release was unspecified.
Visit original source