High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation

Alibaba Group’s Z-Image Team and The Chinese University of Hong Kong present Z-Image Turbo++, a 2-step image generator distilled from the 8-step Z-Image Turbo teacher and built on the 6B-parameter Z-Image S3-DiT foundation. The paper targets the hard limit of few-step diffusion by combining Distribution-Aligned Adversarial Learning, Step-Decoupled Parameterization, and End-to-End Training with Iterative Regularization. The key novelty is to use 8-step teacher-generated images as GAN “real” samples instead of external photos, which the authors report is more stable and yields cleaner outputs; they also split the two denoising steps into independently updated weights, effectively doubling step-specific capacity. In evaluation on OneIGBench, GenEval, DPGBench, and LongTextBench, the final 2-step model reaches 52.50 on OneIGBench, 75.70 on GenEval, 85.86 on DPGBench, and 91.62/89.88 on LongText-CN/LongText-EN, substantially narrowing the gap to the 8-step teacher while preserving sharp texture and stronger text rendering than prior 2-step baselines such as TwinFlow and DMD2. The ablations show that removing the step-1 loss drops OneIGBench to 50.16 and LongText-CN to 84.49, confirming that iterative regularization is critical for keeping the first step meaningful. Training uses 16 H100 GPUs, 20,000 iterations, and about 80 hours, with a memory-efficient backpropagation trick that passes the step-2 gradient into step 1 via an inherited loss.