CyberRealistic Z-Image Turbo by
Cyberdelia (v4.0, from the creator's own
HF mirror), converted from the ComfyUI
single-file checkpoint to the
mflux layout and quantized to
4bit for Apple-silicon Macs. Text encoder (Qwen3-4B), VAE, and tokenizer come from the base
Tongyi-MAI/Z-Image-Turbo (Apache-2.0).
A photorealism-focused finetune of Z-Image-Turbo: ~9 steps, no CFG, multilingual prompts
(Korean included). The 4-bit build runs on a 16 GB Mac.
1from mflux.models.z_image.variants import ZImageTurbo
2m = ZImageTurbo(model_path="avlp12/CyberRealistic-Z-Image-Turbo-v4-mflux-4bit")
3m.generate_image(seed=7, prompt="candid photo of ...", num_inference_steps=9).save(path="out.png")
Used as a built-in model in
Alis Studio.
License: CreativeML OpenRAIL-M (inherited from the finetune; see the use restrictions).
Credit: Cyberdelia (finetune), Tongyi-MAI (base model), mflux (MLX runtime).
