Inpaint

A fully local AI photo inpainting & upscaling desktop app. A C# / Avalonia desktop rewrite of Inpaint-web (the WebGPU/WASM browser version): model inference, image processing, and UI are all C# — no JavaScript, no server, and images never leave your machine.

🖌️ Inpaint (MI-GAN) · 🔍 Upscale ×4 (Real-ESRGAN) · 📦 Export PNG / JPEG / WebP · 🌳 Git-style history · 🔒 100% local & offline

⬇️ Download · 🎬 Demo · ✨ Features · 📥 Models · 🐞 Report an issue · English | 简体中文

🎬 Demo

Inpainting — brush over the content you want to remove and MI-GAN fills it in from the surrounding pixels

Upscaling — Real-ESRGAN tiled upscaling with seam suppression

Settings — theme, language, acceleration device, and more; every change takes effect immediately

✨ Features

• Inpainting: brush over an area and it gets repaired — runs as soon as you release the stroke (switchable back to button mode in Settings).
• Super-resolution (×4 upscaling): Real-ESRGAN tiled upscaling, with CoreML GPU acceleration by default on macOS.
• Fine-grained editing: canvas zoom / pan for masking small details; the mouse wheel adjusts brush size directly over the canvas.
• Generation history: a git-style branching history tree — undo, branch, or go back to the original; the node limit is configurable (25 by default).
• Export & compression: on save, choose PNG / JPEG / WebP and quality with a live output size estimate (the estimated bytes are exactly the bytes written to disk), plus a 1:1 pixel-sampled preview — hold to compare against the original, drag to inspect different regions; JPEG automatically composites transparent areas onto a white background. Encoding uses the built-in Skia, so there are no extra dependencies.
• Personalization: dark / light / follow-system theme, Simplified Chinese / English UI, plus default brush size, history limit, and more — all adjustable in Settings.
• Fully local: models download automatically and are cached on first use, after which everything works offline; the startup update check is opt-in and off by default.

📦 Download

Head to Releases for the package matching your platform (macOS / Windows portable & Setup installer / Linux). The corresponding ONNX model downloads automatically the first time you use inpainting or upscaling.

🛠️ Building from source

Requires the .NET 10 SDK.

dotnet build Inpaint.slnx
dotnet run --project src/Inpaint.App

Unit tests: dotnet test (no model files or network required).

🧱 Project structure

Project	Responsibility
src/Inpaint.Core	Image layout conversions (Bgra8888 ↔ RGB CHW, mask binarization); zero dependencies
src/Inpaint.Inference	ONNX Runtime inference: model download & local caching, MI-GAN inpainting, Real-ESRGAN tiled super-resolution
src/Inpaint.App	Avalonia UI: brush editing canvas, history management, progress & status

📥 Models

Same ONNX models as the web version (migan_pipeline_v2.onnx, realesrgan-x4.onnx). The first time you use a feature, its model downloads automatically from HuggingFace (with a fallback source on failure) and is cached under the app data directory:

• macOS: ~/Library/Application Support/Inpaint/models/
• Windows: %APPDATA%\Inpaint\models\
• Linux: ~/.local/share/Inpaint/models/

If your network is restricted, you can download the models manually and place them in the directory above. When HuggingFace is unreachable directly, go through a proxy via environment variables (https_proxy, etc.), or download from hf-mirror.com and place the files manually; the Settings window can also open the model cache folder directly.

🔑 Key conventions (ported from the web version)

• Model inputs: image [1,3,H,W] uint8 (RGB) + mask [1,1,H,W] uint8; in the mask, 0 = area to inpaint, 255 = keep (white brush strokes are mapped to 0 via grayscale weights, matching the web version's markProcess semantics).
• Super-resolution tiling: 64×64 tiles with 6px overlap padding on all sides, clamped to edge pixels when out of bounds; the 52×52 core region is output at 4× scale.

⚡ GPU acceleration

Differs by engine; the super-resolution acceleration device is selectable in Settings (Auto / CPU / GPU, effective from the next session):

• Super-resolution (Real-ESRGAN, fully convolutional): defaults to CoreML on macOS — measured on an M2, a 64×64 tile drops from 540ms (CPU) to 11ms, at the cost of a one-time 3–4s session compilation.
• Inpainting (MI-GAN): stays on CPU — CoreML can only take over 375 of its 559 nodes, and the partition-copy overhead turns a 0.4s inference into 69s.
• The INPAINT_EP environment variable has the highest priority: cpu forces everything back to CPU, coreml force-enables CoreML (extremely slow on MI-GAN, for experiments only).
• To enable DirectML on Windows: add the Microsoft.ML.OnnxRuntime.DirectML package and append the DML EP in OrtConfig.MakeSessionOptions.

📄 License

MIT