nihui/qwenimage-ncnn-vulkan

ncnn implementation of Qwen-Image-2.1, with text-to-image, image editing, multi-reference image input and transparent RGBA output

What it solves

This project provides a high-performance, portable implementation of the Qwen-Image-2.1 model, allowing users to generate and edit images on a wide range of hardware without requiring heavy dependencies like Python, PyTorch, or CUDA.

How it works

It uses the ncnn universal neural network inference framework and Vulkan for GPU acceleration. This allows the model to run on NVIDIA, AMD, Intel, and Apple Silicon GPUs across Windows, Linux, and macOS. It supports full BF16 precision without quantization and implements automatic memory management (tiling and low-VRAM modes) to fit the model into as little as 2GB of VRAM.

Who it’s for

Users who want to run Qwen-Image-2.1 for image generation and editing on consumer-grade hardware, especially those who prefer standalone executables over complex environment setups.

Highlights

  • Broad Hardware Support: Works on NVIDIA, AMD, Intel, and Apple Silicon GPUs, as well as CPUs.
  • Versatile Image Tasks: Supports text-to-image, image editing, and transparent RGBA output.
  • Advanced Controls: Includes support for ControlNet (pose, canny, etc.), LoRAs, and up to 10 reference images for multi-reference input.
  • Portable: Distributed as a standalone executable with no Python or CUDA dependencies.
  • Optimized Performance: Includes a Turbo version (8-step generation) and supports dynamic output resolutions.

Related

  • Project
    perfectgf/lora-dataset-studio

    A browser-based studio for building, curating, captioning, and training image datasets for LoRAs, integrating ai-toolkit and ComfyUI.

  • Project
    no8d/ComfyUI-NO8D-controls

    ComfyUI‑NO8D‑controls is a set of custom ComfyUI nodes that streamline prompt handling, LoRA stacking, image loading, generation, compositing, and dataset saving. It includes a node that calls an external LLM (OpenAI‑compatible or Ollama) to generate prompts from text or images, plus visual tools for managing prompt libraries and building image grids.

  • Project
    six-nut/PocketMen-with-you

    A local generative tool that transforms two or more photos into consistent, animated character sprites for Codex companions using open-weight image models.

  • Project
    Linfeng-Tang/SwinFusion

    SwinFusion is a Swin Transformer-based framework for general image fusion that combines information from multiple source images across various domains, such as infrared, visible, and medical imaging.