CyberRealistic v9 Z-Image Turbo — GGUF FULL Bundle

🖼️ Main advantages - WHY specifically v9 ? (with examples)

The key difference between Z-Image CyberRealistic v9 and v8 is not merely that it is "another version," but rather a shift in prompting philosophy. V9 moves away from tag stacks toward descriptive language: it better understands natural phrasing about lighting, scene and materials, relies less on booster words like "masterpiece" or "8k", and delivers more natural-looking skin, consistent anatomy and realistic lighting.

Below are six prompts that demonstrate these differences in practice. All examples were generated with the same seed and sampler settings on both v8 and v9 for a fair comparison.

What v9 brings	Example
Natural skin texture and micro-detail	V9 preserves natural skin texture instead of polishing it into a waxy surface: visible pores, forehead lines, crow's feet around the eyes, individual stubble hairs and uneven skin tone all read clearly. Soft window light from camera left reveals the surface relief without flattening it. In v8, such prompts often produced overly smooth, plastic-looking skin where fine detail either disappeared or turned into noise.
Consistent hand anatomy in complex poses	V9 keeps hands anatomically coherent when they interact with objects — here, the left hand frets the strings while the right hand strums near the soundhole. Fingers remain distinct, knuckles are correctly placed, and the grip looks physically plausible. In v8, such scenes often produced fused fingers, extra joints, or unnaturally long nails.
Realistic mixed lighting	V9 handles scenes with two light sources of different colour temperatures — here, cold overhead fluorescent light and warm morning sunlight from a window behind the subject. The face is lit primarily by the overhead fixture, while the warm backlight adds separation from the background. In v8, such mixed lighting often collapsed into a single source or became overly dramatic and unnatural.
Faithful material textures	V9 renders materials through direct description of their physical properties rather than through tags like detailed texture: cracked leather with oil stains and frayed stitching, ceramic with crazing in the glaze, brass with a dull patina and fine scratches. Soft north-facing light from the upper left reveals the surface relief of each object. In v8, such materials were often smoothed into a uniform surface.
Stable composition in wide scenes	In v9, wide scenes with several characters no longer collapse: figures remain distinguishable, faces are not duplicated, limbs are not merged, and the natural arrangement across foreground, midground and background is preserved. In v8, such shots often lost the background plane or merged subjects together.
Complete image from a minimal prompt	V9 produces a finished photographic frame from a short descriptive prompt — no masterpiece, 8k or ultra detailed boosters required. Here, a single sentence about an elderly woman mending a net on a wooden dock at dawn, with fog over the water and natural light, is enough to get a coherent scene with proper composition and detail. In v8, such prompts without quality boosters often resulted in pale or unfinished images.

💡 Tip: These PNGs have the ComfyUI workflow and prompt embedded! Don't just copy the image from the browser preview. Go to the Files and versions tab, open the image there, and click "Download" to get the original file with all metadata intact.

All images generated locally on a GTX 1070 (8 GB VRAM), 1024×1024.

The full ComfyUI workflow used for all examples is available in examples/example_workflow.jpg.

Original checkpoint downloaded from Civitai: CyberRealistic Z-Image Turbo.

Converted for use with ComfyUI and with stable-diffusion.cpp directly.

📦 What's in this repository

Everything you need is in this repo — diffusion model, text encoder and VAE. Nothing to download from other pages.

Diffusion model (Z-Image Turbo, Lumina2 architecture)

Note on quantization naming: Z-Image (Lumina2 architecture) does not benefit from mixed _S/_M/_L quantization recipes — llama-quantize produces byte-identical files to the base _K types. This repository therefore publishes only base quants: Q3_K, Q4_K, Q5_K, Q6_K, Q8_0. The Q4_K_S file is kept alongside Q4_K as a size-compatible alternative.

Text encoder (Qwen3-4B, in clip/)

VAE

Examples

The examples/ folder contains the six reference images shown above (e0.png … e5.png) and the full ComfyUI workflow used to generate them (example_workflow.jpg).

Recommended combo for GTX 1070 (8 GB)

Diffusion Q4_K + text encoder Q4_K_M + VAE — ~8 GB total, fits with room to spare.

For a GTX 1070 (8 GB), the recommended combo is: diffusion Q4_K + text encoder Q4_K_M + VAE — total ~8 GB, fits with room to spare.

Note on quantization naming: Z-Image (Lumina2 architecture) does not benefit from mixed _S/_M/_L quantization recipes — llama-quantize produces byte-identical files to the base _K types. Only Q4_K and Q4_K_M differ. This repository therefore publishes only base quants: Q3_K, Q4_K, Q4_K_M,Q5_K, Q6_K, Q8_0.

Why this file

The original checkpoint is distributed in FP16 (safetensors, ~11.5 GB), which does not fit into 8 GB VRAM. This Q4_K quantization brings the diffusion model down to ~5.2 GB, leaving room for the text encoder and VAE on mid-range GPUs.

What you need besides files from this repo

Abolutely nothing !! This is full-pack bundle

🚀 ComfyUI workflow

The working path is martin-rizzo/ComfyUI-ZImagePowerNodes. It loads Z-Image GGUF directly and applies the correct latent/conditioning shapes.

Required nodes

Install the Power Nodes pack:

bash

cd ComfyUI/custom_nodes

git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes.git

Recommended sampler settings

steps: 4 (Turbo is distilled; anything above ~8 starts to burn)

cfg: 1.0

sampler: euler

turbo_creativity: no

old_scheduler: no

noise_injection: no

alternative_refiner: no

width/height: 1024 (must be multiples of 32)

Alternative: standard ComfyUI GGUF nodes

If you really want to use the built-in UnetLoader (GGUF) instead of Power Nodes, you need city96/ComfyUI-GGUF on PR #392 or later:

bash

cd ComfyUI/custom_nodes/ComfyUI-GGUF

git fetch origin pull/392/head:pr392

git checkout pr392

That fixes the x_pad_token shape mismatch. In practice, though, Power Nodes is more reliable for Z-Image Turbo, and that's what's been verified here.

🙏 Credits

Original model: cyberrealistic-z-image-turbo-catalyst by Cyberdelia on Civitai.

Base architecture: Z-Image / Z-Image Turbo by Tongyi Lab (Alibaba).

ComfyUI nodes: martin-rizzo/ComfyUI-ZImagePowerNodes.

Conversion toolchain: city96/ComfyUI-GGUF + llama.cpp fork.

Runtime: leejet/stable-diffusion.cpp.

Quantization and packaging: gimly13.

📄 License

All components in this repository are distributed under creativeml-openrail-m:

Original diffusion checkpoint (Catalyst) — creativeml-openrail-m.

Qwen3-4B text encoder — Apache 2.0 (Alibaba).

Z-Image VAE (ae.safetensors) — Apache 2.0 (Tongyi Lab).

See the Civitai model page for details on the original model.