onnx-web-optimizer 0.1.0

One-command ONNX model optimizer for web deployment

onnx-web-optimizer

One-command ONNX model optimizer for web deployment.

Runs a three-step pipeline — graph simplification → FP16 quantization → ORT conversion — to produce the smallest, fastest-loading model for use with onnxruntime-web.

---

Why

Loading a raw ONNX model in the browser is slow. This tool automates the steps most teams do manually:

Step	Tool	Typical Result
Graph simplification	onnxsim	Removes redundant nodes
FP16 quantization	onnxconverter-common	~50% size reduction
ORT format conversion	onnxruntime	Faster in-browser parse time

A 168 MB U2Net model becomes ~85 MB after the full pipeline.

---

Installation

pip install onnx-web-optimizer

---

Usage

# Full pipeline (recommended)
onnx-web-opt model.onnx -o ./output

# With verbose output
onnx-web-opt model.onnx -o ./output --verbose

# Skip individual steps
onnx-web-opt model.onnx --skip-simplify
onnx-web-opt model.onnx --skip-fp16
onnx-web-opt model.onnx --skip-ort

# Check version
onnx-web-opt --version

Output

onnx-web-optimizer 0.1.0
Input  : /path/to/model.onnx
Output : /path/to/output
────────────────────────────────────────────────────
[1/3] Simplifying model graph...
✓ Simplification complete: output/model_simplified.onnx
[2/3] Quantizing to FP16...
✓ FP16 quantization complete: output/model_simplified_fp16.onnx
[3/3] Converting to .ort format...
✓ ORT conversion complete: output/model_simplified_fp16.ort
────────────────────────────────────────────────────
✅ Done. Output: output/model_simplified_fp16.ort

---

Loading the Optimized Model in the Browser

import * as ort from 'onnxruntime-web';

const resp = await fetch('https://your-cdn.com/model_simplified_fp16.ort');
const buffer = await resp.arrayBuffer(); // must be ArrayBuffer, not a ReadableStream

const session = await ort.InferenceSession.create(buffer, {
  executionProviders: ['webgpu', 'wasm'],
  graphOptimizationLevel: 'all',
});

---

Requirements

• Python >= 3.8
• onnxruntime >= 1.16.0

---

Real-World Usage

BulkPicTools uses onnx-web-optimizer as part of its AI model deployment pipeline.

BulkPicTools is a privacy-first, browser-based bulk image processing tool — all processing runs locally in the user's browser via WebAssembly. No images are ever uploaded to a server.

The AI-powered features (e.g. Background Remover) rely on ONNX models optimized with this tool before being served to end users.

Optimized Model Registry

Models currently optimized and deployed via this tool in production:

Model	Task	License	Original Source	Optimizations Applied
U2Net	Background Removal	Apache 2.0	xuebinqin/U-2-Net	onnxsim → FP16 → .ort

✅ All models are open-source with MIT or Apache 2.0 licenses.
🔒 Models are served via Cloudflare R2 and run entirely in the user's browser — no data is sent to any server.

---

Who Is This For?

If you are building a browser-based AI tool and need to:

• Reduce model download size for end users
• Speed up model initialization in the browser
• Serve optimized models from a CDN (Cloudflare R2, S3, etc.)

...then this tool is for you.

---

License

MIT © 2026 kbmjj123