StableDfusion

A desktop-friendly web GUI for leejet/stable-diffusion.cpp, modeled on the architecture of its sibling project LLama-GUI. No frameworks, no bundler — just a Python stdlib HTTP server and vanilla JS.

Video generation is still a work in progress, so you may encounter some errors.

What it is

A two-mode launcher for stable-diffusion.cpp:

• Generate (primary): a gallery workflow that runs sd-cli one-shot per image — prompt → live step preview → result thumbnail → history. Covers image, video, upscale, convert, and metadata modes.
• Server & API: launch sd-server persistently and hit its OpenAI (/v1) / SDAPI (/sdapi) / sdcpp (/sdcpp) endpoints, optionally over a Cloudflare tunnel.

Both binaries are downloaded and managed by the Install tab — you don't need a local build of stable-diffusion.cpp.

Quick start

Requirements: Python ≥ 3.11. (Node is optional, only for frontend smoke tests.)

# 1. Install (creates .venv, pip-installs requirements, optional npm deps)
./install.sh                 # macOS / Linux
install-windows.bat          # Windows

# 2. Launch (opens the browser, runs the server)
./start.sh                   # macOS / Linux
start-windows.bat            # Windows

Or run directly:

python server.py             # → http://127.0.0.1:5250

On first launch, open the Install tab to download a stable-diffusion.cpp release and pick a backend for your platform (CUDA 12, Vulkan, ROCm, CPU-AVX2, Metal, …). Then switch to Generate Image and start creating.

Tabs

Tab	Purpose
Install	Download / update / remove sd-cli + sd-server binaries; pick a backend
Generate Image	txt2img / img2img with live preview, gallery, and history
Generate Video	Video (.webm) generation
Upscale	Upscale an image with an upscaler model
Convert	Convert checkpoints to .gguf
Configure	Inspect and tweak every sd-cli / sd-server flag with a live command preview
Server & API	Run sd-server persistently + endpoint docs for OpenAI/SDAPI/sdcpp
HF Download	Browse and stream multi-file model bundles from Hugging Face
Presets	Save / restore / export generation presets

Features

• One-shot generation with per-step progress parsing and live preview polling
• Model bundles drive the file pickers per model type (sd1, sdxl, sd3, flux1/2, qwen_image, wan, ltx, z_image, custom) — each applies sensible defaults for dimensions, steps, CFG, and required components
• Hugging Face downloads auto-routed to models/{diffusion,vae,text-encoders,loras,upscalers}
• Cloudflare tunnel — one-click public URL, cloudflared auto-downloaded
• Git auto-update with safe dirty-path detection and pip install on pull
• Open folder shortcuts to models/, output/, presets/, etc.

How it works

python server.py  →  backend/app.py  (stdlib ThreadingHTTPServer)
                     ├─ serves ui/ as the web root
                     ├─ /api/*  → backend/routes/*.py (33 endpoints)
                     └─ /v1, /sdapi, /sdcpp/*  → proxied to running sd-server (127.0.0.1:1234)

• sd-cli is invoked one-shot per generation (backend/services/generate_service.py).
• sd-server is a separate persistent process with its own lock, so it can run alongside an in-progress generation.
• All UI state lives in one shared window.SDGui.flagCore; writes go through a single setFlagValue() setter, so Generate and Configure always stay in sync.
• ui/js/flags/definitions.js is the single source of truth for every flag, audited against sd-cli -h.

Configuration

Env var	Default	Purpose
SD_GUI_HOST	127.0.0.1	Bind host (0.0.0.0 or * for LAN access)
SD_GUI_PORT	5250	GUI port (distinct from LLama-GUI's 5240)
SD_GUI_ALLOWED_HOSTS	—	Comma-separated extra allowed hosts (LAN IPs / hostnames), admitted as http://<host>:<port> origins
SD_GUI_PROXY_TIMEOUT	1800	Timeout (seconds) per proxied /v1 / /sdapi / /sdcpp request to sd-server

Runtime layout (auto-created on boot): models/, output/ (+ .preview/, .gallery/), presets/, sdcpp/bin/, tools/cloudflared/.

Project layout

server.py            # entrypoint → backend.app.main()
backend/             # app, config, routing, state, http, routes/, services/
ui/                  # index.html + partials/ (tabs), js/ (flags, generate, …), css/
docs/                # directory.md (full reference), cli_flags_report.txt
install.sh / start.sh   (and -windows.bat variants)

For the complete as-built reference, see docs/directory.md.

License

MIT.