Z-Image-Turbo_clear

• Guide (External site): English | Japanese

• Hugging Face

• VAE: Z-Image_clear_vae

Compare to Z-Image-Turbo (Original)

Z-Image-Turbo_clear | Z-Image-Turbo

The Z-Image-Turbo_clear model is a fine-tuned version of Z-Image-Turbo that improves details compared to the original.

For better results, please use it with the dedicated VAE, Z-Image_clear_vae.

Which Precision Should I Use?

+----------+------+----------+----------+--------------------+-----------+
| Format   | Sign | Exponent | Mantissa | Significant Digits | Precision |
+----------+------+----------+----------+--------------------+-----------+
| FP32     | 1bit | 8bit     | 23bit    | 7–8 digits         | ✅        |
| BF16     | 1bit | 8bit     | 7bit     | 2–3 digits         | 👍        |
| FP16     | 1bit | 5bit     | 10bit    | 3–4 digits         | 👍        |
| FP8(e5m2)| 1bit | 5bit     | 2bit     | 1–2 digits         | ⚠️        |
| FP8(e4m3)| 1bit | 4bit     | 3bit     | 1–2 digits         | ⚠️        |
+----------+------+----------+----------+--------------------+-----------+

• Exponent: Stability

• Mantissa: Quality

Anime Illustration Comparison

Hand Close-up

• FP32: Highest precision, high VRAM usage

• BF16: Most popular in Z-Image-Turbo model, good for LoRA compatibility

• FP16: Slightly better precision than BF16, less performance degradation on pre-RTX 2000 or non-NVIDIA GPUs

• FP8(e4m3): Lower precision, low VRAM usage, accelerated on RTX 4000 series and later

If you are unsure which precision to use, it is safest to choose the BF16 format.

Recomended Settings

• Sampler:

  • euler

  • res_multistep

• Scheduler: beta

• Steps: 9

• ModelSamplingAuraFlow: shift 7.5 (3-9)

Upscaler

your choice, examples:

• 4x-SwinIR-L_GAN

• 4x_NMKD-Superscale-SP_178000_G

Tips (External Sites)

• ComfyUI VRAM management: English | Japanese

• Steps and Scheduler: English | Japanese

• Color Correct custom nodes in ComfyUI: English | Japanese

License

• Apache-2.0