README.md

metadata

title: ONNX Runtime Web Image Upscale in Browser
emoji: 🔍
colorFrom: blue
colorTo: indigo
sdk: static
app_file: dist/index.html
pinned: false

ONNX Runtime Web Inspector

Static browser app for loading a local .onnx model and image, running inference with onnxruntime-web, and comparing the latest output to the previous run.

Files

• index.html — Vite entry HTML used for local dev and production builds.
• src/ — Vue 3 app source (script setup components, composables, and shared helpers).
• dist/ — production build output generated by Vite. Hugging Face Static Spaces should serve dist/index.html.
• package.json — Bun scripts and frontend dependencies.
• vite.config.ts — Vite config with Vue and Tailwind 4.
• tsconfig.json — TypeScript config for the Vue app.
• convert_openmodeldb_to_onnx.py — Python script for exporting OpenModelDB-compatible .pth / .safetensors image models to ONNX.
• pyproject.toml — local uv project metadata and pinned Python compatibility.
• uv.lock — locked dependency resolution for reproducible uv sync.
• requirements.txt — exact local dependency snapshot for uv pip install -r.

Run the web app

Install frontend dependencies once:

bun install

Start the local Vite dev server:

bun run dev

Then open the local URL printed by Vite.

To verify the production bundle locally:

bun run build
bun run preview

Hugging Face Static Space

This repository is configured as a static Space and must serve the built Vite output, not the source index.html.

The YAML front matter at the top of this README.md includes:

• app_build_command: bun install && bun run build
• app_file: dist/index.html

That makes Hugging Face:

1. Install the Bun dependencies.
2. Build the app into dist/.
3. Serve dist/index.html, which correctly references the hashed files in dist/assets/.

If you change the frontend build output location, update app_file to match.

Convert a checkpoint to ONNX

Set up a local uv environment first:

uv python install 3.11.11
uv venv --python 3.11.11 --seed
source .venv/bin/activate
uv pip install -r requirements.txt

Or, if you prefer the locked project metadata route:

uv sync

Then run the converter from the local uv environment:

uv run python ./convert_openmodeldb_to_onnx.py /path/to/model.safetensors --output /path/to/model.onnx --seed 42

Useful flags:

• --width / --height control the dummy trace size.
• --device cuda exports on GPU when needed.
• --static-shape disables dynamic height/width axes.
• --no-check skips onnx.checker.

Notes

• The web app currently binds only the first model input and expects it to be a float32 image tensor.
• The Python converter targets OpenModelDB-style image models supported by spandrel / spandrel-extra-arches.