Z-Image Turbo Quality: DyPE Sampling and Resolution Optimization

TL;DR

• Z-Image Turbo is a 6B-parameter distilled model with 8 NFEs that achieves "Very High" visual quality but "Low" diversity at its native 1024×1024 training resolution; for how spectral progression alignment works, see the DyPE Working Mechanism section.
• CFG scale must be <strong>0</strong> for Z-Image Turbo, not "moderate" — higher values cause oversharpening artifacts because Decoupled-DMD already bakes guidance into the student model; for fixes, see the Common Errors and Fixes section.
• DyPE enables native high-res generation through spectral progression alignment, with exponent values tuned per resolution tier; for recommended workflow parameters, see the Configuration Optimization section.
• The distilled model trades sample diversity for speed; the base Z-Image model at 50 steps offers "Medium" diversity and "High" quality, while Omni-Base offers "High" diversity; see the Distilled Model Trade-offs section for details.
• Quick answers to common questions: see the FAQ section.

Sources

• Z-Image Project
• Z-Image Turbo Model Card
• DyPE README
• ComfyUI-DyPE Plugin
• DyPE Paper
• Z-Image Turbo DeJPEG-LoRA
• ComfyUI Reference Workflow JSON

Introduction

Z-Image Turbo achieves remarkable output quality in just <a href="https://github.com/Tongyi-MAI/Z-Image">8 NFEs</a>, making it one of the fastest diffusion transformers available. However, its distilled architecture introduces constraints that are often misunderstood by practitioners. The model cannot simply have its CFG scale increased to improve coherence — doing so double-applies guidance and degrades results. Understanding these trade-offs is essential for anyone generating production-quality images at scale.

This guide covers why CFG behaves differently on Turbo models, how DyPE (Dynamic Position Extrapolation) extends the model beyond its native resolution through frequency-aware positional encoding, and where exactly the distillation process narrows the model's output distribution compared to the base model.

The CFG Correction

The most dangerous misconception circulating about Z-Image Turbo is that a moderate CFG scale produces the best natural results. This is wrong for the Turbo distilled model, though it applies to the base Z-Image at 50 steps.

Z-Image Turbo's Decoupled-DMD distillation merges CFG Augmentation (CA) into the student during training. The model learns to produce CFG-like guidance internally without needing explicit classifier-free guidance at inference. Setting guidance_scale to any positive value on Turbo double-applies this mechanism, leading to oversharpening and artifact amplification. The official model card states guidance_scale=0.0 with the explicit note "Guidance should be 0 for the Turbo models."

The base Z-Image model (non-Turbo), by contrast, uses CFG at 50 steps and benefits from explicit guidance at inference.

How DyPE Works: Spectral Progression Alignment

DyPE is a training-free method enabling diffusion transformers to generate at resolutions far beyond their training data with no additional sampling cost. The technique exploits an inherent spectral ordering in the diffusion process.

The Core Mechanism

Diffusion denoising follows a frequency hierarchy: low-frequency (large-scale structure) converges early in the sampling trajectory, while high-frequency (fine detail) evolves throughout. DyPE makes positional encoding time-aware through a scheduler function κ(t) = λs · t^λt that modulates extrapolation strength at each step.

At the start of sampling (t≈1), this formula yields near-maximal scaling κ(1)=λs, stabilizing global structure through aggressive positional extrapolation. As sampling progresses toward completion (t≈0), the scheduler decays to κ(0)=1, removing extrapolation so the model can access high-frequency perceptual detail exactly when it is needed.

Resolution-Dependent Exponent Tuning

The exponent λt controls how aggressively the extrapolation decays:

The DyPE method vision_yarn is specifically recommended for Z-Image and non-square images at 4K+ because it decouples structure from texture: low frequencies are scaled to fit the canvas aspect ratio while high frequencies are scaled uniformly, preventing the stretching and structural artifacts that static extrapolation methods produce.

Where the Distilled Model Trades Away Detail

Z-Image Turbo's "Low" diversity rating is the most concrete indicator of what distillation sacrifices.

The Diversity Spectrum

These ratings come from the official Z-Image model zoo. Source: Model Zoo

What Changes During Distillation

The 8-step constraint forces the model to take larger denoising steps than the base model's 50-step trajectory. This cannot capture the full multimodal distribution the longer sampling schedule can access.

Specific trade-offs:

• Sample diversity: RL-based DMDR (Distribution Matching Dice with Reinforcement Learning) sharpens outputs toward preferred aesthetics but narrows the space of possible valid outputs for any given prompt
• Rollout consistency: DMD collapse toward high-likelihood modes means the model produces more similar outputs across repeated sampling of the same prompt
• Rare or unusual concepts: Prompts requesting unusual style combinations or rare compositional elements receive less variation across samples

What the model gains: Speed (<a href="https://github.com/Tongyi-MAI/Z-Image">8 NFEs</a> vs. 50 steps), visual fidelity at the modal peak, and native high-resolution generation through DyPE.

Optimal Configuration for Z-Image Turbo + DyPE

The recommended ComfyUI workflow parameters, verified against the published workflow JSON:

Model components:

• Text encoder: qwen_3_4b.safetensors (type: lumina2)
• VAE: ae.safetensors
• Diffusion model: z_image_turbo_bf16.safetensors

Step Budget Analysis

The 9-step budget for Z-Image Turbo is fixed by the architecture -- the model was distilled specifically for this step count. Adding more steps does not improve quality because the student model was not trained to denoise at intermediate timesteps. Similarly, the base Z-Image model at 50 steps and the base model with DyPE at 28 steps each represent their respective training budgets.

VRAM and Throughput

Z-Image Turbo fits within 16 GB VRAM consumer devices at native <a href="https://github.com/Tongyi-MAI/Z-Image">1024×1024</a> resolution. The DyPE node adds negligible overhead since it patches positional embeddings on-the-fly during the forward pass.

Common Errors and Fixes

White Speckle Noise

Symptom: High-frequency grain in focused image areas like hair or eyes at extreme resolutions. Cause: Scaling the model's attention mechanism beyond its training resolution limits produces amplification of high-frequency noise patterns. Fix: Apply the DeJPEG-LoRA at strength 1.0, which smooths high-frequency artifacts while improving proportion accuracy. Source: DeJPEG-LoRA

Proportion Distortion at Extreme Resolutions

Symptom: Geometric stretching or warping of subjects at 4K+ resolutions. Cause: Higher dype_exponent (2.0) maintains strong positional scaling throughout the sampling trajectory, which can over-stabilize spatial relationships when the aspect ratio diverges from training data. Fix: Reduce dype_exponent (e.g., 2.0 → 1.0) at the cost of some noise reduction. Source: ComfyUI-DyPE

Legacy Method Artifacts

Symptom: Stretching and structural artifacts when using yarn or ntk at 4K+. Cause: Static positional extrapolation does not adapt to the spectral progression of the diffusion process, causing misalignment between low-frequency structure stabilization and high-frequency detail recovery. Fix: Switch to vision_yarn, which dynamically balances aspect-ratio handling and frequency scaling internally. Source: ComfyUI-DyPE

FAQ

How many steps does Z-Image Turbo need? 9 steps total (8 DiT forwards). This is the architectural constraint of the 8-NFE distillation. More steps do not improve quality.

What CFG scale is correct for Z-Image Turbo? 0.0 (disabled). The distilled model internally incorporates guidance through Decoupled-DMD CFG Augmentation. Any positive CFG value double-applies this and degrades output quality.

Does DyPE work with Z-Image Turbo? Yes, the ComfyUI-DyPE plugin explicitly supports Z-Image (Lumina 2) models with the model_type=zimage selector. The published reference workflow uses vision_yarn with dype_exponent=2.0 at 2048×1152. Source: ComfyUI-DyPE

Where does Z-Image Turbo lose detail compared to the base model? Primarily in sample diversity. The 8-step distillation cannot capture the full multimodal distribution available through the base model's 50-step rollout. RL-based DMDR further narrows outputs toward a preferred aesthetic mode.

Can I use Z-Image Turbo at resolutions above 2048×1152? Yes, DyPE enables native generation at 4K+ and beyond. The model was not tested beyond 2160×2160 in published workflows, though DyPE was demonstrated at 4096×4096 on related architectures. Source: DyPE Paper

Summary

Z-Image Turbo delivers exceptional visual quality at <a href="https://github.com/Tongyi-MAI/Z-Image">8 NFEs</a> but requires CFG=0 and careful DyPE configuration to avoid artifacts. The distilled model's key trade-off is diversity for quality: where the base Z-Image captures varied interpretations of a prompt, Turbo consistently produces the single most probable output with sharp structural fidelity. At resolutions above native <a href="https://github.com/Tongyi-MAI/Z-Image">1024×1024</a>, DyPE's spectral progression alignment extends the model's capabilities with zero additional sampling cost, though extreme resolutions amplify speckle noise that DeJPEG-LoRA mitigation addresses. For common implementation questions, see the <a href="/internal:z-image-turbo-dype-sampling-resolution-steps-faq">FAQ</a> section.