🎨 SwiftDiffusion — Stable Diffusion GUI

🇵🇱 Polski

Stop wrestling with the command line. Start creating.
A clean, fast, VRAM-friendly GUI for Stable Diffusion 1.5 — built with PyQt6.

---

🖼️ Screenshots

Main generation window

ADetailer — automatic face enhancement

Built-in model downloader for CivitAI & HuggingFace

---

🚀 What can it do?

✍️ Prompting Tools

• Wildcards — use __token__ syntax in your prompt; replaced with a random line from wildcards/token.txt each generation
• Prompt Builder — compose prompts visually from tag categories (quality, style, lighting, artists) with a single click
• Prompt weighting (compel) — fine-tune the influence of words and phrases using standard (word:1.2) syntax, no extra setup required
• Textual Inversion embeddings — drop .pt / .bin / .safetensors files into models/embeddings/; they appear automatically in the Prompt Builder and are ready to use after model load
• Prompt Builder — built-in tag manager with categorized tags (positive and negative), live preview for both, and one-click copy to T2I. Save and restore prompt combinations via history (prompts_history.json, last 20) and favorites (prompts_favorites.json, unlimited). Separate Load/Copy/Delete actions per entry per both tag sets. 🎲 Random button picks configurable tags per positive category.
• Tag search — filter all tag buttons (positive and negative) by typing in a search field. Matching results appear in a dedicated tab with synced toggle state.
• Wildcards tab — browse and toggle wildcards from the wildcards/ folder directly in Prompt Builder. Each .txt file appears as a checkable button; tooltip previews the first 5 lines. Selected wildcards use __name__ syntax and are restored from history/favorites.
• Custom tag categories — drop your own .json file into tags/ with {"label": "My Tags", "tags": ["tag1", "tag2"]} and it appears as a new tab immediately, no restart needed. Built-in categories stay in their original order; your files are added alphabetically below them.
• Style Presets — one-click tag sets in tags/presets/. Each file is one preset: {"name": "My Style", "icon": "✨", "tags": ["tag1", "tag2"]}. Add or remove files anytime — changes appear instantly. Presets are exclusive (selecting one replaces the other).
• Batch generation — generate up to 10 images sequentially with a single click. Each item uses a different seed. Auto upscale is applied per-item. After completion, a scrollable thumbnail bar lets you browse all results and compare raw vs upscaled versions.
• CLIP Interrogator — built-in prompt reverse engineering. Load a CLIP model (ViT-B/32 or ViT-L/14), drop any image, and get a ranked list of matching terms across 8 categories — quality, colors, medium, artist, style, lighting, effects, composition. Custom .json candidate files supported. GPU/CUDA toggle. One-click copy to Text2Image.

🖌️ Generation Modes

Mode	What it's for
Text2Image	Generate from a prompt with full control over sampler, scheduler, CFG, steps, and dimensions
Img2Img	Transform an existing image guided by your prompt
Inpainting	Paint a mask over any area and regenerate just that part — with full Undo/Redo (Ctrl+Z/Y)
ControlNet (Canny)	Guide generation using the edge structure of a reference image
ADetailer	Automatically detect and enhance faces using YOLOv8 — zero extra VRAM cost
Upscaler	High-quality upscaling via the spandrel library
CLIP Interrogator	Analyze any image and reverse-engineer its prompt — detects quality, colors, medium, artist, style, lighting, effects, and composition

⚙️ Smart Settings

• First Launch Wizard — choose your language and theme before anything else
• 7 dark themes — Dark, Amber, Nord, Dracula, Monokai, Forest, Ocean — with optional custom accent color
• Performance controls — VRAM Slicing, Attention Slicing, Tiled VAE, CPU Offloading, auto VRAM clear
• Fully translatable — add any language by dropping a JSON file into locales/

📦 Model Management

• URL Downloader — paste a CivitAI or HuggingFace link and SwiftDiffusion handles the rest: downloads, categorizes, and makes the model available instantly
• Live refresh — add a model or embedding to disk and the UI picks it up automatically, no restart needed (QFileSystemWatcher)
• Latent Mixology Station — blend up to 5 LoRA adapters simultaneously with a visual weight mixer
• Auto-detects .safetensors, .pth, .onnx, and other common formats

🖼️ Gallery & PNG Metadata

• Browse generated images in the built-in file explorer
• One-click recall — read prompt, seed, and all settings from PNG metadata and restore them to the pipeline instantly
• Floating HTML guides for ControlNet and Inpainting

⚠️ Known limitations

• SDXL was evaluated but dropped — from_single_file() crashes on 6 GB GPUs without an error message. SwiftDiffusion is optimized and tested exclusively for SD 1.5.

📊 Live Resource Monitor

Real-time VRAM, RAM, GPU load, and temperature — visible in the sidebar at all times.

---

🛠️ Installation

Windows (recommended)

1. Clone or download this repo
2. Run install.bat
3. Run start.bat

install.bat sets up a virtual environment, installs PyTorch with CUDA 12.8 support, and pulls all dependencies automatically.

Linux

pip install -r requirements.txt
python main.py

---

📋 Requirements

Python	3.12.10
GPU	CUDA-compatible, 6 GB VRAM recommended (GTX 1060 or better)
OS	Windows / Linux (PyQt6 required)

---

📂 Project Structure

SwiftDiffusion/
├── main.py                    # UI skeleton & entry point
├── boot.py                    # CUDA health check on import
├── engine.py                  # Diffusion pipeline logic
├── worker.py                  # Background QThread workers
├── model_manager.py           # Model loading, LoRA, file watchers
├── generation_controller.py   # T2I / Img2Img / Upscale orchestration
├── mode_controllers.py        # Inpaint / ControlNet / ADetailer controllers
├── config.py                  # Settings & i18n loader
├── widgets/                   # UI component package
│   ├── __init__.py
│   ├── dialogs.py             # Settings, gallery, about dialogs
│   ├── inpaint_canvas.py      # Mask editor canvas
│   ├── widgets_common.py      # ClickableLabel, sliders, LoRA items
│   ├── model_downloader.py    # CivitAI / HF downloaders
│   ├── flow_layout.py         # Custom FlowLayout for Prompt Builder
│   ├── prompt_builder.py      # Prompt Builder tab
│   ├── resource_monitor.py    # Live VRAM/RAM monitor widget
│   └── clip_interrogator.py   # CLIP Interrogator tab
├── models_registry.py         # Model scanner & registry
├── url_downloader.py          # Legacy downloader helpers
├── scraper.py                 # Model search scraper
├── install.bat                # Windows installer (bilingual)
├── start.bat                  # Windows launcher
├── locales/                   # Translation files (en, pl, …)
├── wildcards/                 # Wildcard text files
├── tags/                      # Prompt Builder tag categories
└── docs/                      # Floating HTML guides

---

☕ Support

If SwiftDiffusion saves you time, consider buying me a coffee: