CyberRealistic Z-Image Turbo v4.0 — GGUF FULL Bundle
Quantized GGUF version of cyberdelia/cyberrealistic_zimage
Converted for use with ComfyUI and with stable-diffusion.cpp directly.
🖼️ Examples
All generated locally on a GTX 1070 (8 GB VRAM), 1024×1024.
💡 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.
 Fantasy character — 1024×1024, 6 steps |  Wet hair, rain — 1024×1024, 6 steps |
 Interior scene with steam — 1024×1024, 6 steps |  Golden hour portrait — 1024×1024, 6 steps |
 Text + cyberpunk atmosphere — 1024×1024, 6 steps |  Product + text rendering — 1024×1024, 6 steps |
📦 What's in this repository
| Component | File | Size | Notes |
|---|---|---|---|
| Diffusion model (light) | CyberRealistic_zit_v4.0-Q3_K.gguf | 4.31 GB | Q3_K quant — for very tight VRAM budgets |
| Diffusion model (balanced) | CyberRealistic_zit_v4.0-Q4_K.gguf | 5.15 GB | Q4_K quant — recommended for 8 GB VRAM |
| Diffusion model (compact) | CyberRealistic_zit_v4.0-Q4_K_S.gguf | 4.83 GB | Q4_K_S — slightly smaller than Q4_K, good middle ground |
| Diffusion model (quality) | CyberRealistic_zit_v4.0-Q5_K.gguf | 5.68 GB | Q5_K quant — better detail, still fits 8 GB |
| Diffusion model (high) | CyberRealistic_zit_v4.0-Q6_K.gguf | 6.05 GB | Q6_K quant — for 12+ GB cards |
| Diffusion model (max) | CyberRealistic_zit_v4.0-Q8_0.gguf | 7.34 GB | Q8_0 quant — near-lossless, needs 12+ GB |
| Text encoder (light) | clip/Qwen3-4B-Q3_K_M.gguf | 2.08 GB | If you're very tight on VRAM |
| Text encoder (best) | clip/Qwen3-4B-Q4_K_M.gguf | 2.50 GB | Optimal balance for 8 GB cards |
| Text encoder (quality) | clip/Qwen3-4B-Q6_K.gguf | 3.31 GB | For 12+ GB cards |
| VAE | vae/ae.safetensors | 335 MB | 16-channel Z-Image VAE |
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/_Lquantization recipes —llama-quantizeproduces byte-identical files to the base_Ktypes. This repository therefore publishes only base quants:Q3_K,Q4_K,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 ~3.2 GB, leaving room for the text encoder and VAE on mid-range GPUs.
What you need besides this file
Z-Image Turbo is a composite model. This GGUF contains only the diffusion model. To generate images you also need:
- Text encoder:
Qwen3-4B-Q4_K_M.gguf(or the officialqwen_3_4b.safetensors— both work with Power Nodes). - VAE:
ae.safetensorsfrom the original Z-Image release (16-channel VAE).
🚀 ComfyUI workflow
Important: the standard ComfyUI GGUF nodes (UnetLoader (GGUF), CLIPLoader (GGUF)) do not work with Z-Image diffusion models converted via convert.py + llama-quantize. You'll get:
RuntimeError: Error(s) in loading state_dict for NextDiT:
size mismatch for x_pad_token: copying a param with shape torch.Size([3840]) ...
size mismatch for cap_pad_token: copying a param with shape torch.Size([3840]) ...
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:
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: cyberdelia/cyberrealistic_zimage
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
📄 License All components in this repository are distributed under Apache 2.0:
Original diffusion checkpoint (cyberrealistic_zimage) — Apache 2.0
Qwen3-4B text encoder — Apache 2.0 (Alibaba)
Z-Image VAE (ae.safetensors) — Apache 2.0 (Tongyi Lab)
See cyberdelia/cyberrealistic_zimage for details on the original model.