SwitchGen

A Linux-native AI image generator with a GTK4 interface. Uses ComfyUI as an embedded library for Stable Diffusion workflows.

Features

• Text-to-image generation
• Image-to-image transformation
• Inpainting with mask support
• Audio generation (Stable Audio)
• 3D novel view synthesis (Zero123)
• Built-in model downloader for HuggingFace models
• Real-time generation progress
• VRAM monitoring and cleanup

Installation

Arch Linux (AUR)

yay -S switchgen-git

Manual Installation

1. Clone with submodules

git clone --recursive https://github.com/Djwarf/switchgen.git
cd switchgen

If already cloned without submodules:

git submodule update --init --recursive

2. Install system dependencies

# Arch Linux
sudo pacman -S gtk4 libadwaita python-gobject python-pytorch python-torchvision

# Other distributions
# Install GTK4, libadwaita, PyGObject, and PyTorch with CUDA support

3. Install Python package

pip install -e .

Usage

Launch the application:

switchgen

Or run as a module:

python -m switchgen

First Run

1. Click the download button in the header bar to open the model manager
2. Download required models (checkpoints, VAE, CLIP, etc.)
3. Select a workflow type and checkpoint
4. Enter a prompt and click Generate

Project Structure

switchgen/
├── src/switchgen/
│   ├── core/
│   │   ├── comfy_init.py    # ComfyUI initialization
│   │   ├── config.py        # Path and app configuration
│   │   ├── downloader.py    # HuggingFace model downloader
│   │   ├── engine.py        # Generation engine
│   │   ├── models.py        # Model catalog
│   │   ├── queue.py         # Job queue system
│   │   └── workflows.py     # Workflow builders
│   │
│   ├── ui/
│   │   ├── main_window.py   # Main GTK4 window
│   │   ├── model_dialog.py  # Model download dialog
│   │   └── styles/          # CSS themes
│   │
│   └── resources/
│       └── fonts/           # Bundled fonts
│
├── vendor/
│   └── ComfyUI/             # Bundled ComfyUI engine (submodule)
│
├── models/                  # Downloaded models
├── custom_nodes/            # User custom nodes
├── output/                  # Generated images
└── temp/                    # Temporary files

Supported Workflows

Workflow	Description	Required Models
Text2Img	Generate images from text prompts	Checkpoint, VAE
Img2Img	Transform existing images	Checkpoint, VAE
Inpaint	Edit specific regions with masks	Checkpoint, VAE
Audio	Generate audio from text	Stable Audio checkpoint
3D	Generate novel views of objects	Zero123 checkpoint, CLIP ViT-L

Model Storage

Models are stored in the models/ directory:

models/
├── checkpoints/      # Main model files (.safetensors, .ckpt)
├── vae/              # VAE models
├── clip_vision/      # CLIP vision encoders
├── text_encoders/    # Text encoders (T5, etc.)
├── controlnet/       # ControlNet models
├── loras/            # LoRA adapters
└── upscale_models/   # Upscaling models

Requirements

• Python 3.10 or later
• GTK4 and libadwaita
• PyTorch with CUDA support
• NVIDIA GPU with 8GB+ VRAM recommended

Configuration

The application auto-detects paths. To override:

• Set COMFYUI_PATH environment variable to use an external ComfyUI installation
• Models are stored in ./models/ relative to the application root

Troubleshooting

Missing models error

Open the model download dialog and download the required models for your workflow.

VRAM out of memory

• Close other GPU applications
• Use smaller image dimensions
• Try a different checkpoint (some use less VRAM)

Custom nodes not loading

Place custom node folders in the custom_nodes/ directory and restart the application.

License

MIT License. See LICENSE file for details.

Credits

• ComfyUI by comfyanonymous
• Stable Diffusion by Stability AI
• GTK4 and libadwaita by GNOME