Tech 3 min read

BEYOND_REALITY_Z_IMAGE - a photorealistic people-focused model based on Z-Image Turbo

IkesanContents

While looking at Z-Image derivative models, I noticed a checkpoint called BEYOND_REALITY_Z_IMAGE on ModelScope. It is a fine-tune specialized for photorealistic people, with a film-photography-like texture.

Model overview

ItemDetails
AuthorNurburgring (Zhang Chi)
BaseZ-Image Turbo
MethodLoRA training on a people dataset, then merged
LicenseApache License 2.0
Versionsv1.0 to v3.0
DownloadsAbout 3,800

ModelScope: Nurburgring/BEYOND_REALITY_Z_IMAGE

Features

  • Skin texture optimization: More detailed skin texture on people
  • Film-photography aesthetics: Color and lighting feel closer to film cameras
  • Environmental detail: Backgrounds and props are rendered more carefully
  • Keeps Z-Image Turbo’s speed: Can generate in 10 to 15 steps
SettingValue
SamplerEuler + Simple
Steps10-15
CFG1-2

Because it is based on Z-Image Turbo, you can still get high-quality results with fewer steps. A lower CFG is enough too.

Will it run on an M1 Max 64GB?

Conclusion: yes.

Z-Image Turbo memory requirements

ComponentSize
Z-Image Turbo (BF16)About 12 GB
Qwen3 4B text encoderAbout 7 GB
VAEA few hundred MB
TotalAbout 20 GB

An M1 Max with 64 GB of unified memory has plenty of room. A quantized version is even lighter.

QuantizationSize
BF16 (full)About 12 GB
FP8About 6 GB
Q4_K_M (GGUF)About 4 GB

Running on Apple Silicon

It works with ComfyUI + Metal support. If there is an MLX version, it could be optimized even further. There are also reports that stable-diffusion.cpp can run the GGUF-quantized version with only 4 GB of VRAM.

Where Z-Image derivatives fit

The Z-Image family is roughly organized like this:

ModelUse
Z-ImageBase model. Supports LoRA and ControlNet
Z-Image-TurboDistilled version. Fast generation in 8 steps
Z-Image-Omni-BaseMultimodal base model
Z-Image-EditImage-editing specialization

BEYOND_REALITY_Z_IMAGE is a checkpoint that merges a photorealistic people-focused LoRA on top of Z-Image Turbo. It keeps Turbo’s speed while improving the quality of portraits.

Z-Image and image-generation articles on this blog:

References