Tag
This paper introduces Z-Image Turbo++, a two-step image generation model distilled from an eight-step teacher using distribution-aligned adversarial learning, step-decoupled parameterization, and end-to-end training with iterative regularization to narrow the quality gap with multi-step generation.