Tongyi-MAI

/

Z-Image-Turbo

like 4.11k

Follow

Tongyi-MAI 3.86k

Hi Z-Image community,

I'd like to share a project: https://github.com/nihui/zimage-ncnn-vulkan — a pure C++ implementation of Z-Image inference using https://github.com/Tencent/ncnn and Vulkan GPU acceleration.

What is it?

This is a standalone, cross-platform implementation of Z-Image image generation that runs entirely through ncnn + Vulkan, with no dependency on Python, PyTorch, or CUDA runtime.

Key Highlights

Cross-platform & Cross-vendor GPU support

• Runs on Windows, Linux, and macOS
• Supports Intel, AMD, NVIDIA, and Apple Silicon GPUs via Vulkan
• Also supports CPU-only mode (-g -1)

Zero runtime dependency

• Download the executable + model files, and you're ready to go
• No need to install Python, pip, PyTorch, CUDA toolkit, or any other environment
• Truly portable — just copy and run

Lightweight & efficient native implementation

• Written in C/C++, compiled to a single binary
• All dependencies (libwebp, libjpeg-turbo, libpng, etc.) are built-in
• Supports both z-image and z-image-turbo models

Low barrier to entry

• Minimum: 16GB system RAM + any Vulkan-capable GPU (Linux/macOS)
• On Windows: (half of system RAM) + GPU VRAM >= 16GB
• Recommended: 32GB RAM + 16GB VRAM GPU with tensor core support

Usage

# Basic usage
zimage-ncnn-vulkan -p "a wandering astronaut on mars, photo realistic, 8k" -o output.png

# Use turbo model with custom resolution
zimage-ncnn-vulkan -m z-image-turbo -p "your prompt" -W 1024 -H 1024 -o output.png

# Specify GPU device
zimage-ncnn-vulkan -p "your prompt" -g 0 -o output.png

Why this matters for Z-Image adoption

1. Lowers the barrier — Users don't need to set up a Python ML environment. Download, place models, run.
2. Broader hardware support — Not limited to NVIDIA CUDA. AMD, Intel, and Apple Silicon users can participate.
3. Easier distribution — A single portable executable makes it simple to integrate into other applications or workflows.
4. Demonstrates model portability — Shows that Z-Image's architecture can be efficiently deployed outside the PyTorch ecosystem.

Build from source

git clone https://github.com/nihui/zimage-ncnn-vulkan.git
cd zimage-ncnn-vulkan
git submodule update --init --recursive --depth 1
mkdir build && cd build
cmake ../src
cmake --build . -j 4

Model files are available on https://huggingface.co/nihui-szyl/z-image-ncnn/tree/main.

Links

• GitHub: https://github.com/nihui/zimage-ncnn-vulkan
• Models: https://huggingface.co/nihui-szyl/z-image-ncnn/tree/main
• ncnn framework: https://github.com/Tencent/ncnn

Feedback and contributions are welcome!