Z-Image Turbo

Highlights

• 8 function evaluations (NFE) to an image; the official example writes num_inference_steps=9, which is effectively 8 forward passes
• 6B parameters plus Apache-2.0, the lightest and cleanest combination on this list
• A full Z-Image exists alongside it; Turbo is the distilled, accelerated variant
• Mature ecosystem: the count of ZIT base-model LoRA on Civitai is already substantial

Scope

• Good for: fast everyday generation, batch drafts
• Good for: low-VRAM machines and cases that need VRAM left over for LoRA
• Not for: commercial final renders that require native 2K

Tips

• Keep CFG at 1; on distilled models like this, adjusting CFG mostly just makes results worse
• At 8 steps, keep LoRA weight at or below 1.0 — overexposure is easy
• To gain detail, switch to the full Z-Image rather than adding steps

Cautions

• Turbo and full-version weights are not interchangeable, and their templates differ as well

Parameters

Steps

8 (example shows 9)

CFG

1 (distilled model; CFG is not the way to add detail)

Resolution

1024×1024

Text encoder

Qwen3-VL family (see official template)

Sampling

Built into the official template; changing steps yields limited gain

Deployment

16GB verdict

The most worry-free option: 6B weights fit with room left over for LoRA and upscaling.

VRAM notes

z_image_turbo_bf16 and several ZIT merge models are already on your machine, so you can compare directly.

Suggested starting point

8 steps, CFG 1, 1024²; for larger images, generate at 1024 first and then run an upscale model.

Comparison

6B + Apache-2.0 + 8 steps is the best answer in the low-VRAM, clean-license slot, and roughly 650K downloads in 30 days bears that out. The weak spots are in-image text and native 2K: those two jobs belong to Qwen-Image 2.1 or Ideogram 4.