Z-Image Turbo++ generates high-fidelity images in just 2 steps

A team of researchers led by Dongyang Liu has unveiled Z-Image Turbo++, a breakthrough in few-step diffusion distillation that delivers high-fidelity images in just two denoising steps. While previous methods achieved decent quality in 4-8 steps, pushing further to 2 steps suffered from increased task difficulty and limited model capacity. The team tackles this with three tailored design choices that bring 2-step quality remarkably close to 8-step performance.

The first innovation, Distribution-Aligned Adversarial Learning, replaces external real images with teacher-generated images as adversarial targets, making the GAN training more attainable and informative. Second, Step-Decoupled Parameterization assigns separate model parameters to each of the two denoising steps, addressing their distinct capacity demands. Third, End-to-End Training with Iterative Regularization lets the first step receive gradients from final image quality while preserving meaningful intermediate outputs via an explicit step-1 loss. Together, these techniques cut required inference steps by 75% versus the teacher model, with minimal quality degradation—paving the way for real-time image generation in production environments.