Z-Image

Z-Image Base vs Turbo: Which One is Better for Your LoRA Training?

4월 28, 2026

Z-Image Base vs Turbo: Which One is Better for Your LoRA Training?

There are two versions of Z-Image out there: Base and Turbo.

If you're new to Z-Image, you probably have one question: which one should I pick? On HuggingFace, Turbo has 1.21 million downloads — 34 times more than Base. But if you want to do LoRA fine-tuning, the answer isn't that simple.

The Short Answer

  • Need LoRA fine-tuning / flexible control → Base
  • Just want fast, high-quality image generation → Turbo

Let's dig deeper.

Side-by-Side Comparison

Feature Z-Image (Base) Z-Image-Turbo
Inference Steps 28~50 8
CFG Support ✅ ❌
Negative Prompts ✅ ❌
Fine-tunable (LoRA) ✅ ❌
Generation Quality High Very High
Diversity High Low
RL Training ❌ ✅

Why Turbo Has 34x More Downloads Than Base

Turbo's core selling points are speed and quality.

How Much Faster?

Turbo needs only 8 inference steps to generate an image, while Base needs 28-50. This means:

  • Base: 50-step inference, ~10-15 seconds per generation (H800)
  • Turbo: 8-step inference, ~1-2 seconds per generation (H800)

On consumer GPUs, Turbo runs comfortably in 16GB VRAM, making the speed advantage even more noticeable.

Why the Quality is Higher

Turbo introduces two key technologies:

  1. Decoupled-DMD: Decoupled Distribution Matching Distillation — splits the two independent mechanisms of traditional distillation algorithms and optimizes them separately, significantly improving few-step generation quality
  2. DMDR (DMD + RL): Fuses Distribution Matching Distillation with Reinforcement Learning during post-training, improving semantic alignment, aesthetic quality, structural consistency, and high-frequency detail richness

In short, Turbo is a "distilled + RL-enhanced" upgrade of Base, achieving better results with fewer steps.

So What's Base For?

If you only generate images, Turbo is the clear choice. But Base has three capabilities Turbo lacks:

1. LoRA Fine-tuning (The Key Difference)

This is Base's biggest killer feature.

Base is a full, undistilled Diffusion Transformer — it can be fine-tuned with LoRA directly. Want to train your own portrait style, product style, or character customization? Base is the way to go.

Turbo has gone through distillation and RL training, so it does not support traditional LoRA fine-tuning. Community attempts to fine-tune LoRA on Turbo have produced unsatisfactory results.

2. CFG and Negative Prompts

Base supports CFG (Classifier-Free Guidance) and negative prompts:

# Base supports this
image = pipe(
    prompt="a cat",
    negative_prompt="blurry, deformed, low quality",  # ← Turbo doesn't support this
    guidance_scale=4,                                   # ← Turbo is fixed at 0
).images[0]

This means you can:

  • Use negative prompts to exclude unwanted elements (blur, extra fingers, deformities)
  • Adjust guidance_scale to control prompt adherence

Turbo's guidance_scale is fixed at 0 — no adjustment possible.

3. Higher Generation Diversity

Base produces more diverse results. With the same prompt, Base generates richer variations in composition and details each time. Turbo is more "consistent" — but also more "monotonous."

For creative work, diversity means more opportunities for inspiration.

Real-World Scenario Guide

Choose Base When:

  • ✅ LoRA Training: Portrait customization, style transfer, character training
  • ✅ Need Precise Control: Use negative prompts to exclude specific elements
  • ✅ Batch Creation Needs Variety: Poster design, product image variation
  • ✅ Research Purposes: Need CFG and full model architecture

Choose Turbo When:

  • ✅ Fast Generation: Idea validation, rapid prototyping
  • ✅ Cost-Sensitive Inference: API calls, commercial deployment
  • ✅ No Fine-tuning Needed: Use the official model directly
  • ✅ Quality First: Highest quality for individual images

My Recommendation

If you're doing LoRA training, don't overthink it — go with Base. Turbo may be fast, but the lack of fine-tuning support is a hard limitation.

If you just generate images, Turbo offers better value. 8-step inference + higher quality + lower VRAM requirements — there's no reason not to choose it.

There's also a hybrid approach: Use Base for fine-tuning, train your LoRA weights, then generate with Base + LoRA. This gives you both flexibility and quality.

FAQ

Q: Will Turbo support LoRA fine-tuning in the future?

The official team hasn't announced any plans. Turbo's distilled architecture is inherently unsuitable for traditional LoRA — future fine-tuning may require new methods.

Q: How much VRAM do I need for LoRA training on Base?

Base is a 6B parameter model. You'll need at least 24GB VRAM (RTX 3090/4090) for smooth training. With LoRA rank 16, VRAM usage is about 12-14GB.

Q: 50 inference steps on Base is too slow. What can I do?

You can drop to 28 steps with minimal quality loss. Or use CPU offloading to reduce VRAM pressure.

Summary

Need Recommendation
LoRA Fine-tuning Base
Fast Generation Turbo
Fine Control (CFG / Negative Prompts) Base
Generation Diversity Base
Single-Image Quality Turbo
Inference Cost Turbo

Neither choice is universally right or wrong — it depends on your core needs.

Z-Image Team

Z-Image Team