Stable Diffusion Studio

The best .NET-native local AI image studio Create, organize, and generate AI images with the control and privacy of running models on your own machine.

---

Overview

Stable Diffusion Studio is a local-first, web-based image generation application built on .NET 10. It provides a polished, fast, and highly configurable user experience for discovering models, organizing projects, composing prompts, generating images, and iterating on creative workflows — without requiring third-party GUI servers like Automatic1111 or ComfyUI.

Key Features

• Image generation — txt2img, img2img, and inpainting with SD 1.5, SDXL, and Flux models via StableDiffusion.NET (CUDA + Vulkan GPU acceleration)
• Model management — Browse and download from HuggingFace and CivitAI, scan local folders, auto-detect model types (Checkpoint, VAE, LoRA, Embedding, ControlNet)
• A1111-style workspace — Model/VAE/LoRA selection, sampler/scheduler/steps/CFG/seed controls, CLIP skip, batch generation, resolution presets
• Project-based workflow — Organize work into projects with persistent generation history
• Presets system — Save/load generation parameter presets, filtered by model and family
• Prompt tools — Prompt history with search, negative prompt quick-insert tags, Ctrl+Enter to generate
• Image gallery — Favorites, NSFW shield with configurable threshold, per-image actions (copy seed/prompt, delete)
• Content safety — Fully-local NSFW detection via NsfwSpy ML.NET, blur/reveal/block modes, A1111-compatible PNG metadata embedding
• Live preview — Skeleton progress card with step counter during generation, SignalR real-time updates
• Background jobs — Non-blocking generation/downloads with progress tracking, cancellation, interactive job detail view
• Performance settings — VRAM presets, thread count, flash attention, VAE tiling, CPU/GPU offloading controls
• Warm industrial design — Custom dark/light theme with Inter + JetBrains Mono typography
• Local-first privacy — All data stays on your machine, no cloud accounts required
• Aspire-powered — Full observability dashboard with OpenTelemetry

GPU Support

CUDA is auto-selected for NVIDIA GPUs. Vulkan is the fallback for universal GPU support. CPU offloading is automatically disabled on CUDA to prevent AVX512 crashes on hybrid Intel CPUs.

---

Quick Start

Prerequisites

• .NET 10 SDK
• .NET Aspire workload: dotnet workload install aspire
• NVIDIA GPU recommended (RTX 3000+ for best performance)

Run the app

git clone https://github.com/JerrettDavis/StableDiffusionStudio.git
cd StableDiffusionStudio

# Direct launch (recommended — fastest startup, most stable)
dotnet run --project src/StableDiffusionStudio.Web
# Open http://localhost:5190

# Or with Aspire dashboard (adds OpenTelemetry monitoring)
dotnet run --project src/StableDiffusionStudio.AppHost

First steps

1. Go to Settings → Storage Roots and add your model directory (e.g., G:\sd.webui\webui\models\Stable-diffusion)
2. Tag additional folders for LoRA, VAE, etc.
3. Go to Models and click Scan Now
4. Go to Generate or create a Project, select a model, enter a prompt, and generate

Run the tests

# Unit and integration tests (819+)
dotnet test --filter "FullyQualifiedName!~E2E"

# E2E BDD tests (Playwright + Reqnroll)
pwsh tests/StableDiffusionStudio.E2E.Tests/bin/Debug/net10.0/playwright.ps1 install chromium
dotnet test tests/StableDiffusionStudio.E2E.Tests

---

Architecture

Built as a modular monolith with strict layer boundaries:

Presentation (Blazor Server + MudBlazor + SignalR)
    ↓
Application (Services, Commands, DTOs, Validation)
    ↓
Domain (Entities, Value Objects, Domain Services)
    ↑
Infrastructure (EF Core + SQLite, StableDiffusion.NET, NsfwSpy, File I/O)

See docs/architecture/README.md for the full architecture overview.

Solution Structure

---

Milestones

• Milestone 1+ — App shell, project CRUD, model scanning, background jobs, settings, dark theme
• Milestone 2 — Unified model provider system (HuggingFace, CivitAI), model downloads, tagged model folders, model type classification
• Milestone 3 — Generation pipeline (txt2img), StableDiffusion.NET inference backend (CUDA + Vulkan), generation workspace with A1111-style controls, gallery with metadata
• Milestone 4 — Presets, prompt history, image favorites, PNG metadata embedding, negative prompt helpers, advanced parameters (CLIP skip, batch count), data management
• Milestone 5 — img2img, inpainting, plugin architecture (ControlNet deferred — requires preprocessor models)

What's New

• img2img — Upload an init image, adjust denoising strength, generate variations
• Inpainting — Paint a mask over an init image to selectively regenerate regions (canvas with configurable brush, touch support)
• Settings export/import — Export all settings to JSON, import from file to restore configuration
• Plugin architecture — IPlugin, IPostProcessor, IModelProviderPlugin interfaces with assembly-scanning PluginManager

Additional features built

• Content safety (NSFW detection + shield toggle)
• Live generation preview (skeleton card with step progress)
• Backend performance settings (VRAM presets, offloading controls)
• Custom design system (warm industrial palette, Inter + JetBrains Mono)
• Flux model support (auto-detect VAE, CLIP-L, T5-XXL components)
• Background generation (survives page navigation)
• Interactive Jobs page with detail view
• Keyboard shortcuts (Ctrl+Enter)
• Recent models quick-switch
• Database health check and diagnostics
• Robust schema migration and repair

---

Contributing

Contributions are welcome! Please see our Contributing Guide for details.

1. Fork the repository
2. Create a feature branch (git checkout -b feat/amazing-feature)
3. Commit your changes using Conventional Commits
4. Push to the branch (git push origin feat/amazing-feature)
5. Open a Pull Request

---

Documentation

• Architecture Overview
• Product Specification
• ADR Decision Records

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Acknowledgments

Built with .NET, MudBlazor, .NET Aspire, StableDiffusion.NET, and NsfwSpy.