Dongyang Liu, Ruoyi Du, David Liu, Dengyang Jiang, Liangchen Li, Qilong Wu, Zhen Li, Steven C.H. Hoi, Hongsheng Li, Peng Gao
5 min
Abstract
Few-step diffusion distillation has become increasingly mature for 4-8-step generation, yet pushing further to 2 steps remains challenging. In this work, we introduce Z-Image Turbo++, a high-quality 2-step image generation model distilled from the 8-step Z-Image Turbo teacher. Our method addresses the central bottlenecks of increased task difficulty and limited model capacity in 2-step generation through three simple but effective design choices tailored to this regime. First, we propose Distribution-Aligned Adversarial Learning, which uses teacher-generated images rather than external real images as real samples for GAN training, providing a more attainable and informative adversarial target. Second, we adopt Step-Decoupled Parameterization, assigning independent model parameters to the two denoising steps to better match their distinct capacity demands. Third, we perform End-to-End Training with Iterative Regularization, allowing the first step to receive gradients from final image quality while preserving a meaningful intermediate generation through an explicit step-1 loss. Together, these designs substantially narrow the quality gap between 2-step and 8-step generation in both qualitative and quantitative evaluations, highlighting the potential of carefully tailored distillation strategies for improving the quality-efficiency trade-off in few-step generation.
Sam: The second innovation addresses a different issue. Normally, a model uses the same internal settings—the same learned parameters—for every step of the process. But step one and step two are doing completely different jobs. Step one is about establishing the basic structure of the image: what objects are present, roughly where they are. Step two is about refinement: sharpening edges, filling in texture, getting the details right.
Alex: And using the same settings for both would be like using a sledgehammer for both demolition and surgery.
Sam: That's a fair analogy. So they use what they call Step-Decoupled Parameterization—each step gets its own independent set of learned weights, its own specialised "brain," if you like. Step one focuses on the "what," and step two focuses on the "how."
Alex: Does giving each step its own separate brain create a problem? That sounds like it would double the storage requirements.
Sam: It does increase the model's storage size, which is a genuine trade-off. The paper acknowledges this. However, in practice, you can distribute the two steps across different processors to keep the overall speed high. The researchers also pair this with a technique called Classifier-Free Guidance—essentially a way of running two versions of the output simultaneously and comparing them to steer the result toward what was requested—which helps maintain image quality without requiring additional steps.
Alex: And the third piece was end-to-end training, right? Making sure the two steps actually cooperate?
Sam: Yes. Without that, you could end up with a situation where step one produces something that's technically fine on its own, but makes step two's job much harder. End-to-end training lets the model learn how the choices made in step one ripple through to the final result, so the two steps coordinate rather than work against each other. They also found that keeping a specific regularisation penalty applied to the first step is important—it prevents the model from taking shortcuts that look good in isolation but degrade the final image.
Alex: So it's not just about raw speed. The order and structure of the thinking still matters, even when you've compressed it down to two steps.
Sam: That's the key finding, really. And the results suggest this combination—the reachable teacher target, the specialised per-step weights, and the coordinated training—closes most of the quality gap between two-step and eight-step generation.
Alex: Are there still things it struggles with?
Sam: Yes. The paper notes it remains less reliable for images containing dense text or highly complex scenes. Those cases require a level of fine-grained precision that two steps, however well-designed, can't fully match. So there's still a genuine trade-off between responsiveness and perfect accuracy in difficult cases.
Alex: That's a reasonable place to land. Instant generation for most tasks, with the understanding that some edge cases still need more time.
Sam: It's a notable step toward making generative AI feel as fluid and immediate as more traditional creative tools—without pretending the hard problems have all been solved.
Alex: Thanks for listening to ResearchPod.