Vibecoded.

Desktop app for local AI texture upscaling on NVIDIA GPUs, with browsing and download from OpenModelDB.

Built with Electron + a Python Spandrel / PyTorch worker (CUDA).

NVIDIA GPU required. This app targets CUDA on an NVIDIA graphics card. AMD / Intel GPUs are not supported. Without a working NVIDIA driver + CUDA-capable GPU, upscaling is not the intended use case.

• NVIDIA CUDA Spandrel inference (tiled, VRAM-friendly)
• Size modes
  • Factor — ×1 / ×2 / ×3 / ×4 / ×6 / ×8 / ×16
  • Longest side — presets 512 → 16384 px, plus a custom value (64–16384)
• Tile size — 128 / 256 / 384 / 512 / 1024 (larger = faster on high VRAM, more memory)
• Export — PNG, TGA, TIFF, DDS (uncompressed ARGB), EXR, WebP, JPEG, BMP
• Alpha preserved — RGB through the model, alpha resized separately
• Before / after compare viewer (drag / paste / load textures)
• OpenModelDB catalog — search, scale filter, pagination, previews, download (GitHub, Google Drive, MEGA, …)
• Models library — import / manage local weights (.pth, .pt, .ckpt, .safetensors)
• Setup tab — thin installers; AI runtime (~4.5 GB CUDA) on first use; optional reuse of an existing ComfyUI / Stability Matrix venv

No Python install is required for the Setup / Portable builds — the app creates or reuses a venv itself.

From Releases:

Both builds are thin: they do not ship the multi‑GB PyTorch stack inside the EXE.

Run a newer *-Setup.exe over an existing install — no manual uninstall.

• Same appId → Windows finds the previous install and overwrites program files in that folder (no second copy).
• AppData (AI runtime, settings, imported models) is kept.
• Shortcuts are refreshed. Close the app before upgrading if it’s running.
• Do not change appId in future builds or upgrades break.

Portable: replace the old *-Portable.exe (and keep OpenModelDB-Upscaler-Data\ next to it).

Open the Setup tab (see Setup tab below). Other tabs stay locked until the AI runtime is ready.

1. Drop / paste / load an image (PNG, TGA, TIFF, JPG, WEBP, BMP, …)
2. Pick a model (or download one from OpenModelDB)
3. Choose Factor or Longest side (preset or custom)
4. Set export format + tile size
5. Upscale → Show file when done

This is the AI environment screen (Spandrel / PyTorch / NVIDIA CUDA). Wheels download to Downloads\OpenModelDB-Upscaler-Cache; a full CUDA env is typically ~4.5 GB.

Until Setup reports ready, Upscale, Models, and OpenModelDB stay locked.

No existing env selected. Click Set up runtime:

1. App creates an isolated venv under its data folder (AppData or portable data dir)
2. Downloads CUDA PyTorch + Spandrel into that venv
3. Clears the download cache when install succeeds

This never modifies ComfyUI, Stability Matrix, or any other Python install.

If you already have torch + CUDA elsewhere (common with ComfyUI / Stability Matrix):

1. Under Existing environment, click Browse venv…
2. Select the venv folder (must contain Scripts\python.exe), e.g.
   …\Data\Packages\ComfyUI\venv
3. The app inspects it and shows whether torch / spandrel / Pillow (and CUDA) are present
4. Choose Install missing packages into:

5. Click Set up runtime (button text updates: Create AppData venv… / Install missing into selected env / Re-check / repair… depending on state)
6. Clear removes the selected path from app settings (does not delete the venv on disk)

Rescan re-runs diagnosis. Open data folder / Open cache jump to the app data dir and the Downloads cache.

Cancel install stops downloads and removes the incomplete AppData venv + download cache. It does not delete or roll back a browsed third-party env.

1. UPSCALER_PYTHON (if set and ready)
2. Selected custom env (if ready)
3. App-managed AppData / portable venv (if ready)
4. Project python/.venv (dev)
5. Otherwise: create AppData venv (or install into selected env if that target is chosen)

Shared download cache (both): %USERPROFILE%\Downloads\OpenModelDB-Upscaler-Cache\

Portable does not write the final runtime into system AppData (only the shared Downloads cache).

Windows uninstall (Setup build) asks whether to also remove AppData + the download cache.

• Yes — deletes app data (app-managed venv, settings, imported models) and the cache
• No — keeps them for a faster reinstall

A browsed / selected existing env is never touched by uninstall — not deleted, not cleaned, not rolled back. Only the app’s own AppData / cache paths can be removed.

git clone https://github.com/MiloAchille/OpenModelDB-Upscaler.git
cd OpenModelDB-Upscaler
npm install
npm run setup
npm start

npm run setup creates python/.venv and installs PyTorch + Spandrel (CUDA on NVIDIA).

You can also leave the project venv empty and use the in-app Setup tab (Browse venv… or AppData install), or set UPSCALER_PYTHON to a specific python.exe.

Weights are not in git (GitHub’s 100 MB limit). Put a file in models/, e.g.:

• 4xLSDIRDAT → models/4xLSDIRDAT.pth

Or use OpenModelDB / Models → Import inside the app. See models/README.md.

build.bat

Produces thin artifacts in dist/:

• OpenModelDB Upscaler-1.0.0-Setup.exe
• OpenModelDB Upscaler-1.0.0-Portable.exe

AI runtime is downloaded on first use via the Setup tab, not during build.bat.

assets/                 # icons, logo, NSIS wizard bitmaps
build/installer.nsh     # uninstall prompt (AppData / cache only)
models/                 # local weights (binaries gitignored)
python/upscale.py       # Spandrel CUDA worker
python/requirements.txt
scripts/                # setup, portable, NSIS helpers
src/main.js             # Electron main
src/runtime-bootstrap.js
src/preload.js
src/renderer/           # UI (Upscale · Models · OpenModelDB · Setup)
build.bat               # Windows Setup + Portable

MIT — see LICENSE.

Model weights you download keep their own licenses (often listed on OpenModelDB).