nihui / qwenimage-ncnn-vulkan

# Qwen-Image-2.1 ncnn Vulkan
:exclamation: :exclamation: :exclamation: More features and optimizations in progress
ncnn implementation of Qwen-Image-2.1, with text-to-image, image editing, multi-reference image input and transparent RGBA output.
qwenimage-ncnn-vulkan uses ncnn project as the universal neural network inference framework.
## Features
- Full **BF16** model precision, no quantization
- Works with **2GB** VRAM, no CPU offload
- **NVIDIA / AMD / Intel / Apple Silicon**
- **Windows / Linux / macOS**
- **CPU** inference is also supported
- **No** CUDA / PyTorch / Python dependency
- **Portable** standalone executable
- Text-to-image
- Image **editing**
- Up to **10** reference images
- **Transparent** RGBA image generation
- **Dynamic** output resolution
- **Batch** generation
- Qwen-Image-2.1-Fun-Acc-LoRAs (**4steps**)
- Qwen-Image-2.1-Turbo (**8steps**)
- **ControlNet** (pose, canny, etc.)
RX9060XT qwem-image-2.1 text-to-image 1024x1024
|dit transformer step|(s/it) less is better|
|:-:|:-:|
|torch-nightly (rocm10.0.0) 20260923|6.12|
|stable-diffusion.cpp (vulkan) 20260927|5.31|
|qwenimage-ncnn-vulkan 20260928|4.17|
## Download
Download Windows/Linux/macOS Executable for Intel/AMD/NVIDIA/Apple-Silicon GPU
**https://github.com/nihui/qwenimage-ncnn-vulkan/releases**
This package includes all the binaries required. It is portable, so no CUDA, PyTorch or Python runtime environment is needed :)
### prepare model files
Download the base qwenimage21 model folder.
For Qwen-Image-2.1-Turbo, also download the qwenimage21-turbo model folder. It contains the Turbo Transformer `.param` and `.bin` files and reuses the base model's other components.
Place the folders side by side under `models/`, or select a model folder with `-m`. Keep the base folder named `qwenimage21` next to the Turbo folder:
 [code] 
Both models support text-to-image and image editing, dynamic output sizes and up to ten reference images. Use `-m models/qwenimage21-turbo` to select Turbo and its automatic eight-step configuration.
Files in the Turbo folder take precedence; missing files are loaded automatically from the sibling `qwenimage21` folder, including the text encoder, vision encoder, VAE, tokenizer, vision position embeddings and default ControlNet. No copying of shared components is needed.
`.param` and `.bin` files are resolved separately, so the Turbo folder can also contain only its three Transformer `.bin` files when the base model's `.param` files are available.
## About Qwen-Image-2.1
Qwen-Image-2.1: A unified text-to-image generation and image editing model open-sourced by the Qwen team. With a 7B parameter visual component across 32 Single-Stream DiT layers, it balances generation quality, inference efficiency, and versatility.
https://github.com/QwenLM/Qwen-Image-2.1
## Usages
### requirements
- Minimum (Linux / macOS): 16GB RAM, any Vulkan capable GPU
- Minimum (Windows):
 *Due to WDDM limitations: Vulkan applications can only use half of the system RAM.*
 *Th