holgt/z-image-burn

Z-Image Turbo (Burn)

Z-Image Turbo model weights converted to Burn's Burnpack format for use with the Burn deep learning framework.

Model Description

Z-Image is a fast image generation model based on flow-matching diffusion, designed for generating high-quality images from text prompts. This repository contains the model weights in Burnpack (.bpk) format, optimized for inference on Apple Silicon (Metal) and other Burn backends.

Files

File	Size	Description
z_image_turbo_bf16.bpk	11 GB	Transformer weights in BF16 precision

Additional Required Files

To run Z-Image, you also need:

• Tokenizer: From holgt/qwen3-0.6b-burn (tokenizer.json)
• Text Encoder: From Comfy-Org/z_image_turbo (qwen_3_4b.safetensors)
• Autoencoder: From black-forest-labs/FLUX.1-dev (ae.safetensors)

Usage

With the z-image-app

See the z-image-app for a macOS GUI application.

Rust Code Example

use burn::backend::candle::{Candle, CandleDevice};
use half::bf16;
use qwen3_burn::{Qwen3Config, Qwen3Model, Qwen3Tokenizer};
use z_image::{GenerateFromTextOpts, modules::ae::AutoEncoderConfig, modules::transformer::ZImageModelConfig};

type Backend = Candle<bf16, i64>;

fn main() {
    let device = CandleDevice::metal(0);
    let model_dir = PathBuf::from("./models");

    // Load components
    let tokenizer = Qwen3Tokenizer::from_file(model_dir.join("qwen3-tokenizer.json")).unwrap();

    let mut text_encoder: Qwen3Model<Backend> = Qwen3Config::z_image_text_encoder().init(&device);
    text_encoder.load_weights(model_dir.join("qwen3_4b_text_encoder.safetensors")).unwrap();

    let mut transformer = ZImageModelConfig::default().init(&device);
    transformer.load_weights(model_dir.join("z_image_turbo_bf16.bpk")).unwrap();

    let mut ae = AutoEncoderConfig::flux_ae().init(&device);
    ae.load_weights(model_dir.join("ae.safetensors")).unwrap();

    // Generate image
    let opts = GenerateFromTextOpts {
        prompt: "A beautiful sunset over mountains".to_string(),
        out_path: PathBuf::from("output.png"),
        width: 512,
        height: 512,
    };

    z_image::generate_from_text(&opts, &tokenizer, &text_encoder, &ae, &transformer, &device).unwrap();
}

Requirements

• Apple Silicon Mac with Metal support, or
• CUDA-capable GPU (with appropriate Burn backend)
• ~16GB RAM for 512x512 images
• Rust 2024 edition

License

Apache 2.0

Acknowledgments

• Burn - Deep learning framework
• Comfy-Org - Original Z-Image model
• black-forest-labs - FLUX autoencoder