See it work

Three steps on one picture, each about 3–10 seconds on an RTX 5070 Ti (7–15 seconds on an M5 Max), straight from the app:

Why MosImage

• 🔒 Private by design. Images, prompts and history stay on your disk. No account and no image uploads; models and optional adapters download from Hugging Face on first use.
• ⚡ Fast. FLUX.2 Klein in NVFP4 runs a 1024 × 1024 image in about 3 seconds on a 16 GB card, and in BF16 in about 7 seconds on an M5 Max.
• 🎯 Precise. Edits stay inside your selection, and every result lands as a layer you can mask, move or undo.
• 🧩 A real editor. Layers, revision history, proposals to compare, and brush, lasso, box and click selection.

Features

Create

• Generate an image from a prompt. In an open document the result is added as a new layer.
• References: put people, pets or objects from your own photos into a new scene ("Anna and Tom riding a roller coaster"), up to three at once.
• Sketch rough coloured strokes and describe them. MosImage renders just the sketch, or the whole picture around it.

Edit

• Replace or Remove a selected area. Pixels outside the selection stay exact.
• Edit the whole picture or a single layer with an instruction, such as "make it a snowy night".
• Fix faces: pick the detected faces to sharpen. Faithful keeps expression and lighting closer; Redraw restores very blurry faces and can change their expression. A reference can guide likeness.
• Enhance adds fine detail to the whole picture or a selection, with Subtle and Strong settings. Large areas render in tiles at their original resolution; small details can change.
• Expand the canvas in any direction, with format presets.
• Crop resizes the canvas without AI, with ratios and exact dimensions. Use C for Crop and Shift+C for Expand.

Select

• Object: click to select with SAM 2.1 Tiny, Alt+click (⌥-click on a Mac) to exclude.
• Find: select by description, such as "the dog" or "hat, beer glass".
• Subject and Remove background, with soft edges for hair and fur.
• Brush, eraser, lasso and box tools, plus grow, shrink, feather and invert.

Work with layers

• Every AI result can land on its own layer, with transform, mask, opacity and order.
• Select several layers, merge them down or flatten, as in Photoshop; the picture itself is a layer too.
• Documents keep a full revision history: undo and redo cover every layer.
• Results come as one to four proposals. Compare them with the original, then apply or discard.

Requirements

Other NVIDIA GPUs can use FLUX.2 Klein 4B (BF16) under Generation settings → Advanced controls. A Mac runs the BF16 weights the NVFP4 builds come from, or quantizes them to 8 or 4 bits in MLX: from 64 GB of memory it uses BF16, from 48 GB 8-bit, below that 4-bit (MOSIMAGE_APPLE_WEIGHTS overrides this). Below 36 GB the engine warms up Fast, the editor starts on Fast, and the VAE works in 512 px tiles to save memory.

Quick start

From the repository root, on Windows:

.\start.cmd

or on a Mac:

./start.sh

The first run creates backend/.venv, installs PyTorch (the CUDA build on Windows, the Metal build on a Mac) and the backend requirements, creates backend/.env and installs the frontend dependencies. It then starts:

• the engine on http://127.0.0.1:8000
• the editor on http://localhost:5173

Models download from Hugging Face on first use. FLUX.2 Klein 9B, the Quality model, is gated: accept its licence on its model page and log in once with hf auth login, or set MOSIMAGE_HF_TOKEN in backend/.env. The other models need no token. On later runs, .\start.cmd -NoInstall or ./start.sh --no-install skips the install checks.

For manual setup, troubleshooting a CPU-only PyTorch install and every runtime option, see the guide.

Models

FLUX.2 Klein does all the AI work: generate, edit, replace, remove, sketch, fix faces, enhance and expand. Enhance on Quality 9B adds its non-commercial detail-enhancer adapter; on Fast it redraws without it. The CPU models handle selection. With the model loaded on the target GPU, a 1024 × 1024 generation takes about 3 s, a full-frame edit about 8 s, and widening a 1254 px photo by half about 10 s. Enhance takes about ten seconds per tile. On an M5 Max, Quality takes about 7 s to generate, 11 s to replace a selection and 15 s for a full-frame edit; Fast about half that. The engine warms up the Quality model in the background at start-up; downloads and loading add time. See the model comparison and Enhance evaluation for measurements.

Desktop app

cd frontend
npm run tauri:build

This builds frontend/src-tauri/target/release/mosimage.exe plus MSI and NSIS installers. The app starts and stops the engine itself, or reuses one that start.cmd already started. Close a running mosimage.exe before rebuilding, because Windows locks the file.

On a Mac, after ./start.sh has set up the engine once:

cd frontend
npm run tauri:build:mac

This builds MosImage.app and installs it in /Applications, so it opens from Launchpad, Spotlight or the Dock like any other app.

Project layout

backend/
  app/
    routes/     API routers: generation, selection, jobs, documents, references
    models/     Model providers behind one ImageModelProvider interface
    ...         Job queue, compositing, selection, documents, storage
  tests/        unittest suite
  tools/        Benchmark and fixture scripts
frontend/
  src/
    hooks/      Engine, jobs, picture, selection, layers, references
    generation/ Submission and command logic
    canvas/     Editor canvas, masks, sketch, expansion
    components/ Panels, dialogs, tool rail, library
  src-tauri/    Desktop shell
benchmarks/     Versioned fixtures and benchmark reports
docs/           Guide, architecture, roadmap and research notes
data/           Local database and images (not in Git)
models/         Optional local checkpoints (not in Git)

Development

# Backend tests and syntax check (on a Mac: .venv/bin/python)
cd backend
.\.venv\Scripts\python.exe -m unittest discover -s tests -t .
.\.venv\Scripts\python.exe -m compileall app

# Frontend tests, type check and build
cd ..\frontend
npm test
npm run build

Contributor conventions live in AGENTS.md. For how the pieces fit together, read the architecture overview.

Documentation

License

The MosImage source code is released under the MIT License.

Model weights are not part of this repository. They download from Hugging Face under their own licences: FLUX.2 Klein 9B is non-commercial, while FLUX.2 Klein 4B, SAM 2.1, Grounding DINO and ViTMatte permit commercial use. Check each model card before you use results commercially.

The screenshots and example images in docs/images were generated with MosImage itself.