andrevp/Z-Image-Turbo-MLX-8bit

Z-Image-Turbo — MLX (8-bit Quantized)

MLX conversion of Tongyi-MAI/Z-Image-Turbo for Apple Silicon.

This is the 8-bit quantized MLX conversion. Linear layer weights are quantized to 8-bit with group_size=64. VAE remains in float16 to preserve image quality.

Model size: 11.37 GB

All Available MLX Variants

About Z-Image-Turbo

Z-Image is an efficient 6B-parameter image generation foundation model using a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture. Z-Image-Turbo is the distilled variant with only 8 NFEs (Number of Function Evaluations), achieving sub-second inference latency.

Key Features

• —Photorealistic image generation with state-of-the-art quality
• —Bilingual text rendering (English & Chinese)
• —Strong instruction adherence
• —8-step inference — distilled via Decoupled-DMD + Reinforcement Learning (DMDR)
• —No CFG required — guidance_scale=0.0

Architecture

The S3-DiT architecture concatenates text tokens, visual semantic tokens, and image VAE tokens at the sequence level as a unified input stream, maximizing parameter efficiency compared to dual-stream approaches.

Quantization Details

Only 2D weight tensors from Linear layers are quantized. Normalization layers, biases, embeddings, and position encodings remain in float16.

Component Sizes

Original Model

• —Source: Tongyi-MAI/Z-Image-Turbo
• —Authors: Tongyi MAI Team (Alibaba)
• —License: Apache 2.0
• —Papers:
• —Z-Image: arXiv:2511.22699
• —Decoupled-DMD: arXiv:2511.22677
• —DMDR: arXiv:2511.13649

Original Usage (PyTorch/CUDA)

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

prompt = "Young Chinese woman in red Hanfu, intricate embroidery, ancient temple backdrop"

image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=9,   # Results in 8 DiT forwards
    guidance_scale=0.0,      # No CFG for Turbo models
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

image.save("example.png")

Conversion Details

• —Converted using MLX {0.30.6} on Apple Silicon
• —Weights converted from bfloat16 to float16
• —SafeTensors format (MLX-compatible)
• —All weight keys preserved and verified
• —VAE kept at float16 across all quantization levels
• —Verified: no NaN/Inf values, all shapes consistent, all index files valid

Citation

@article{z-image2025,
    title={Z-Image: An Efficient Image Generation Foundation Model with Scalable Single Stream Diffusion Transformer},
    author={Tongyi MAI Team},
    journal={arXiv preprint arXiv:2511.22699},
    year={2025}
}

@article{decoupled-dmd2025,
    title={Decoupled Consistency Model Distillation},
    author={Liu et al.},
    journal={arXiv preprint arXiv:2511.22677},
    year={2025}
}

@article{dmdr2025,
    title={DMDR: Fusing DMD with Reinforcement Learning},
    author={Jiang et al.},
    journal={arXiv preprint arXiv:2511.13649},
    year={2025}
}