Z-Image Turbo++: 2-Step Image Generation

• Z-Image Turbo++ is a 2-step image generation model distilled from an 8-step teacher, designed to produce high-fidelity images efficiently.
• It employs distribution-aligned adversarial learning, step-decoupled parameterization, and end-to-end training with iterative regularization to overcome extreme denoising challenges.
• Empirical benchmarks demonstrate near-teacher performance, with notable gains in dense text rendering and complex scene composition despite extra parameter storage.

Searching arXiv for the relevant papers on Z-Image Turbo++ and the Z-Image family. Z-Image Turbo++ is a high-fidelity 2-step image generation model in the Z-Image family, introduced as a student distilled from the 8-step Z-Image-Turbo teacher in order to push few-step diffusion generation from the already mature 4–8-step regime into an extreme 2-step regime without a large collapse in visual quality or instruction following (Liu et al., 10 Jun 2026). Within the broader family, it inherits the single-stream diffusion-transformer lineage of Z-Image, whose base model is a 6.15B-parameter foundation model built on the S3-DiT architecture, while redefining the student parameterization and post-training procedure specifically for ultra-low-NFE generation (Team et al., 27 Nov 2025).

1. Family lineage and model identity

Z-Image Turbo++ sits in a three-level genealogy. Z-Image is the base text-to-image foundation model; Z-Image-Turbo is the accelerated few-step variant produced by few-step distillation plus reward post-training and operating at 8 NFEs; Z-Image Turbo++ is then a further distillation of that turbo teacher into 2 denoising steps (Team et al., 27 Nov 2025, Liu et al., 10 Jun 2026).

This placement matters because Turbo++ is not described as a direct compression of the original long-chain model. The teacher is already a few-step generator, and the paper treats the production of the 8-step Z-Image Turbo teacher as a solved prerequisite. A recurrent misconception is that Turbo++ is simply the turbo model run at a smaller step count. The paper explicitly rejects that interpretation: it is not “the 8-step model run at 2 steps,” and it is not a generic few-step baseline. Instead, it is a 2-step student specifically redesigned for this extreme regime (Liu et al., 10 Jun 2026).

The architectural lineage comes from the Z-Image family’s Scalable Single-Stream Diffusion Transformer (S3-DiT). In that design, text tokens, image VAE latent tokens, and optionally image semantic tokens are concatenated into one unified sequence and processed by the same backbone, rather than by a dual-stream architecture. The base S3-DiT configuration reported for Z-Image has 6.15B total parameters, 30 layers, hidden dimension 3840, 32 attention heads, FFN intermediate dimension 10240, and RoPE dimensions $(d_t, d_h, d_w) = (32, 48, 48)$ (Team et al., 27 Nov 2025).

2. Problem setting: why 2-step generation is different

The central claim of Z-Image Turbo++ is that 2-step generation is qualitatively harder than ordinary few-step distillation. In the usual 4–8-step setting, each denoising step still covers a relatively local segment of the teacher trajectory. In 2-step generation, each step must absorb a much broader transformation: the first step must map near-pure noise to a meaningful intermediate latent, and the second must convert that intermediate into a clean, realistic, prompt-aligned image (Liu et al., 10 Jun 2026).

The paper identifies two distinct bottlenecks. The first is optimization difficulty: if the student is trained against a target distribution that is too far away, the training signal becomes dominated by discrepancies the student cannot realistically close. The second is capacity under step specialization: the two denoising steps perform sharply different functions, so forcing both through one shared parameter set creates interference and under-allocation of capacity (Liu et al., 10 Jun 2026).

Formally, the work remains in the flow matching framework used by Z-Image. The noised sample is written as

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$

where $x_1 \sim p_{\text{data}}$ is the clean image and $\epsilon \sim \mathcal{N}(0,I)$ is noise. A network $v_\phi(x_t,t)$ predicts the corresponding velocity field. This suggests that Turbo++ is best understood not as a new generative paradigm, but as a teacher-aligned compression of an existing flow-matching generator into an unusually aggressive sampling budget (Liu et al., 10 Jun 2026).

3. Three defining design choices

The model is organized around three mechanisms: Distribution-Aligned Adversarial Learning, Step-Decoupled Parameterization, and End-to-End Training with Iterative Regularization (Liu et al., 10 Jun 2026).

Component	Function	Stated purpose
Distribution-Aligned Adversarial Learning	Uses teacher-generated images as GAN real samples	Makes the adversarial target more attainable and informative
Step-Decoupled Parameterization	Uses independent parameters for the two denoising steps	Addresses step specialization and capacity bottlenecks
End-to-End Training with Iterative Regularization	Backpropagates final-image quality through the two-step chain while keeping a step-1 loss	Coordinates both steps while preserving a meaningful intermediate

Distribution-Aligned Adversarial Learning replaces the standard GAN practice of treating external real images as positive samples. Instead, the discriminator uses images generated offline by the 8-step teacher as the “real” distribution. The stated rationale is that, in the 2-step regime, the student’s attainable distribution is much closer to the teacher distribution than to the natural-image distribution. This changes the adversarial problem from “match real photos” to “match a strong 8-step teacher,” which the paper describes as a more attainable and informative objective. The reported effect is both improved training stability and cleaner outputs; with external real photos, the discriminator finds easy cues the student cannot remove, and training becomes less useful (Liu et al., 10 Jun 2026).

Step-Decoupled Parameterization assigns independent model parameters to the two denoising steps. Both copies are initialized from the same 8-step teacher weights, but they are then optimized separately. The paper interprets the benefit as twofold: decoupling reduces multi-task interference between the steps, and it increases effective capacity for a regime in which the first and second denoising maps are highly specialized. A weaker alternative using a shared backbone with per-step LoRA improves some easier benchmarks, but remains markedly worse on more demanding text-oriented metrics, which the authors take as evidence that lightweight adaptation is insufficient in this setting (Liu et al., 10 Jun 2026).

End-to-End Training with Iterative Regularization leverages the fact that full end-to-end optimization across a 2-step chain is feasible. Gradients from final image quality are allowed to flow back into step 1, so the first step is optimized not only for a plausible intermediate but for the final generation objective. However, the paper also reports that end-to-end training alone is inadequate: removing the explicit step-1 loss causes strong degradation, because the model collapses toward an opaque intermediate representation that does not respect the pretrained diffusion model’s iterative denoising prior. The added step-1 supervision therefore acts as an iterative regularizer (Liu et al., 10 Jun 2026).

4. Training objective, equations, and optimization procedure

The overall generator objective is presented as

$$\mathcal{L} = \mathcal{L}_{\text{GAN}} + \lambda \mathcal{L}_{\text{DMD}}.$$

Here the adversarial term supplies the main perceptual alignment pressure, while Distribution Matching Distillation (DMD) provides guidance augmentation and regularization (Liu et al., 10 Jun 2026).

For DMD, the paper reviews the integral KL objective

$$\mathcal{L}_{\text{IKL}}(p_{\text{real}}, p_{\text{fake}}) = \int_0^1 \mathbb{KL}(p_{\text{real},\tau} \| p_{\text{fake},\tau})\, d\tau,$$

and adopts the decoupled interpretation in which the score difference decomposes into a CFG-augmentation term and a distribution-matching term. The constrained re-noising schedule is

$\tau_{\text{CA}} = \tau_{\text{DM}} > t.$

The paper explicitly notes that it does not use the Decoupled-Hybrid schedule, because the GAN term already supplies anti-artifact regularization (Liu et al., 10 Jun 2026).

The adversarial term is the standard GAN objective,

$$\mathcal{L}_{\text{GAN}} = \mathbb{E}_{x \sim p_{\text{real}}}[\log D(x)] + \mathbb{E}_{x \sim p_{\text{fake}}}[\log(1-D(x))],$$

but with a redefined “real” distribution: teacher-generated images rather than external real photographs. This is the key conceptual shift of the method (Liu et al., 10 Jun 2026).

The appendix-level procedural equations make the two-step generator explicit. If the two models are $f_{\theta_0}$ and $$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$0, then

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$1

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$2

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$3

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$4

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$5

with $$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$6. Both the step-1 and step-2 outputs receive GAN and DMD supervision. To preserve end-to-end information flow without holding the full graph in memory, the paper introduces an inherit loss based on the detached gradient from the second step,

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$7

and the actual step-0 objective becomes

$$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$8

with $$x_t = t \cdot x_1 + (1 - t) \cdot \epsilon, \quad t \in [0,1],$$9 in practice (Liu et al., 10 Jun 2026).

The reported optimization setup is highly specific: Adam, learning rate $x_1 \sim p_{\text{data}}$0, $x_1 \sim p_{\text{data}}$1, $x_1 \sim p_{\text{data}}$2, no weight decay, generator-to-guidance-model update ratio 1:5, 20,000 iterations, EMA decay 0.99, 16 H100 GPUs, global batch size 64, and about 80 hours of training time. The generator DMD weight is $x_1 \sim p_{\text{data}}$3, the guidance-model diffusion loss weight is $x_1 \sim p_{\text{data}}$4, the GAN generator and discriminator weights are each $x_1 \sim p_{\text{data}}$5, and the inherit-loss weight is 0.1 (Liu et al., 10 Jun 2026).

5. Empirical performance and benchmark position

The headline empirical claim is that the final 2-step student narrows the gap to the 8-step Z-Image-Turbo teacher to an unusual degree. The reported scores are 52.50 OneIG, 75.70 GenEval, 85.86 DPG-Bench, 91.62 LongText-CN, and 89.88 LongText-EN, versus the teacher’s 52.84, 75.01, 84.86, 92.56, and 91.74, respectively (Liu et al., 10 Jun 2026).

This comparison is notable because the 2-step student is sometimes slightly below the teacher and sometimes slightly above it, depending on the benchmark. The paper is careful, however, not to claim perfect parity. It states that a small gap remains, especially for dense text rendering and complex scenes with secondary objects, even if aggregate metrics show near-teacher performance (Liu et al., 10 Jun 2026).

Against other 2-step baselines, the reported results are consistently strong. Compared with Twin Flow, the method improves from 51.38 to 52.50 on OneIG and from 78.65 / 71.99 to 91.62 / 89.88 on LongText-CN / LongText-EN. Compared with the paper’s DMD2 reimplementation, it improves from 50.70 to 52.50 on OneIG and from 85.30 / 80.99 to 91.62 / 89.88 on LongText-CN / LongText-EN, although DMD2 retains a slightly higher GenEval score (76.12 vs. 75.70) (Liu et al., 10 Jun 2026).

A particularly important negative control is Z-Image-Turbo run at 2 steps, which scores 50.94 OneIG, 76.53 GenEval, 85.78 DPG-Bench, 78.48 LongText-CN, and 72.89 LongText-EN. This directly supports the paper’s claim that simple truncation is insufficient. The strongest gains of Turbo++ appear on the more discriminative LongText and instruction-following benchmarks, which the authors treat as the clearest tests of genuine 2-step compression quality (Liu et al., 10 Jun 2026).

6. Ablations, limitations, and significance

The ablation results are central to the paper’s interpretation of its own contributions. Using teacher-generated images as GAN real samples improves stability and removes artifacts, even though the benchmark movement is mixed in isolation. Step-decoupled parameterization yields large gains over both shared-weight and per-step LoRA alternatives: the shared-weight variant falls to 50.67 OneIG, 73.62 GenEval, 85.51 DPG-Bench, 87.64 LongText-CN, and 81.14 LongText-EN; the per-step LoRA variant reaches 50.96, 76.10, 86.58, 80.71, and 76.90. Adding end-to-end training lifts performance further, and removing the step-1 loss causes a sharp decline to 50.16 OneIG, 71.02 GenEval, 83.98 DPG-Bench, 84.49 LongText-CN, and 82.07 LongText-EN (Liu et al., 10 Jun 2026).

The main limitation explicitly acknowledged is extra parameter storage. Because the two denoising steps use independent weights, the generator effectively doubles in size relative to a shared-weight student. The paper argues that this may be manageable in server settings through pipelining, but it is a real cost for memory-constrained deployment (Liu et al., 10 Jun 2026).

A second limitation is that the remaining failure modes are not random; they concentrate in dense text rendering, secondary objects, and complex scenes. This suggests that even when the global metric gap to the teacher becomes small, the hardest compositional and typography-heavy cases continue to expose the weakness of the ultra-few-step regime (Liu et al., 10 Jun 2026).

In the context of the broader Z-Image program, Turbo++ extends the efficiency trajectory begun by Z-Image-Turbo, which had already compressed the base model from roughly 100 NFEs with classifier-free guidance to 8 NFEs while claiming sub-second inference latency on an H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM) (Team et al., 27 Nov 2025). Turbo++ therefore represents a further shift from “fast few-step generation” to “minimal-step generation,” with its significance lying less in a new backbone than in a new answer to the question of how to distill a modern diffusion-transformer family into just two denoising evaluations (Liu et al., 10 Jun 2026).