Qwen

/

Qwen-Image-2.1

Like 2.01k

Follow

Qwen 106k

Hi Qwen team and everyone,

I have open-sourced qwenimage-ncnn-vulkan, an ncnn/Vulkan implementation of Qwen-Image-2.1:

https://github.com/nihui/qwenimage-ncnn-vulkan

Converted ncnn models are available here:

https://huggingface.co/nihui-szyl/qwen-image-ncnn

The goal of this project is to make Qwen-Image-2.1 easy to run locally on a wide range of consumer hardware, without requiring CUDA, PyTorch, or a Python runtime.

Highlights

• Full BF16 model precision — no weight quantization required

• Works with as little as 2 GB of dedicated GPU memory, with model data backed by system memory, no CPU offload, no harddisk offload

• Cross-platform Vulkan GPU inference

  • NVIDIA
  • AMD
  • Intel
  • Apple Silicon

• Windows / Linux / macOS

• CPU inference is also supported

• No CUDA dependency

• No PyTorch dependency

• No Python runtime required

• Portable standalone executable

• Text-to-image generation

• Image editing

• Up to 10 reference images

• Transparent RGBA image generation

• Dynamic output resolution

• Batch generation

One converted model package supports both text-to-image and image editing.

Example:

qwenimage-ncnn-vulkan \
    -p "A small red kite over a quiet lake." \
    -s 1024,1024 \
    -l 40 \
    -o output.png

Image editing:

qwenimage-ncnn-vulkan \
    -i input.png \
    -p "Change the clothes to a blue jacket." \
    -o output.png

The project is still in an early stage, but the main Qwen-Image-2.1 inference pipeline is already working across different GPU vendors and operating systems.

Feedback, testing results, and contributions are very welcome :)

Source code:
https://github.com/nihui/qwenimage-ncnn-vulkan

ncnn models:
https://huggingface.co/nihui-szyl/qwen-image-ncnn