Alibaba's Qwen-Image-2.0 doubles compression and cuts generation steps from 40 to 4

Alibaba detailed Qwen-Image-2.0, an image model built for much better efficiency without giving up quality.

The model doubles image compression to 16x spatial downsampling, removes the VAE discriminator, and revises the transformer to avoid unstable activations.

It also adds a prompt-expansion module that turns short prompts into richer descriptions, improving text handling and generation quality.

Alibaba says these changes cut generation steps from 40 to 4, which could greatly reduce compute cost and latency.