https://github.com/zdisket/windiffusion

Windows desktop (C++/Qt) app for local Stable Diffusion inference
https://github.com/zdisket/windiffusion

image-generation-ai onnx onnxruntime-directml qt stable-diffusion

Last synced: 30 days ago
JSON representation

Windows desktop (C++/Qt) app for local Stable Diffusion inference

• Host: GitHub
• URL: https://github.com/zdisket/windiffusion
• Owner: ZDisket
• License: mit
• Created: 2024-03-10T01:08:55.000Z (8 months ago)
• Default Branch: master
• Last Pushed: 2024-04-23T06:56:55.000Z (7 months ago)
• Last Synced: 2024-04-24T06:47:07.771Z (7 months ago)
• Topics: image-generation-ai, onnx, onnxruntime-directml, qt, stable-diffusion
• Language: C++
• Homepage:
• Size: 4.69 MB
• Stars: 4
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        


    

        

    



# WinDiffusion

![](assets/demoreel.png)

*Made with RealVisXL Lightning, 8 steps Euler A and upscaled with 4xUltraSharp*

## What is this?

![](assets/porsches.png)

WinDiffusion is a Stable Diffusion frontend written in C++/Qt, without a single line of Python involved, using the ONNX runtime and DirectML to execute models

![](assets/canvasmountain.gif)

*Showcase of combined drawing-img2img Canvas tab*

### So, what's the deal?

1. **Natively compatible with all GPU vendors**. The DirectML backend supports any DirectX 12-capable GPU

2. **Lightweight**. Everything needed to run the models is ~200MB, compared to the around 10GB of pip or conda-installed libraries.

3. **Easy to install**. Installation is a breeze—simply unzip and launch the executable. It's so simple, even your grandma could do it.

4. **Self-contained, reliable**. Without having to lug around lots of libraries, it remains unaffected by unforeseen changes in dependencies.

## Support

Marked with ❌ means not currently available, but is on high priority.

### Supported models (tested)

- ✔️ Stable Diffusion 1.5

- ✔️ Stable Diffusion XL

- ✔️ Stable Diffusion XL Turbo

- ✔️ Stable Diffusion XL Lightning

### Samplers

- ✔️ DPM 2M++ Karras

- ✔️ Euler Ancestral

- ❌ DPM++ SDE Karras (for models that demand it, use Euler Ancestral instead for now)

### Features

- ✔️ Text-to-image

- ✔️ Image-to-image

- ✔️ Inpainting

- ✔️ Upscaling with ESRGAN

- ✔️ (Prompt:1.5) ((weighting))

- ❌ Long prompts (longer than CLIP limit)

- ❌ Face fix

## Compilation

TODO: fill out this section

## Externals (and thanks)

- [Axodox-machinelearning](https://github.com/axodox/axodox-machinelearning): C++ implementation of Stable Diffusion

- [QGoodWindow](https://github.com/antonypro/QGoodWindow): Fancy windows for Qt

- [JSON](https://github.com/nlohmann/json)

- [Qt Color Widgets](https://gitlab.com/mattbas/Qt-Color-Widgets)