bollscoasts/pix2act-weblinx-base-onnx

Pix2Act-WebLINX-Base ONNX

ONNX export of McGill-NLP/pix2act-base-weblinx optimized for browser and server-side inference.

Model Details

Property	Value
Original Model	McGill-NLP/pix2act-base-weblinx
Architecture	Pix2Struct (Vision2Seq)
Parameters	~282M
Export Format	ONNX (opset 17)
Quantization	INT8 dynamic
Model Size	~1.17 GB (quantized)
Browser Eligible	Borderline (works with WebGPU)

Description

Pix2Act-WebLINX is fine-tuned on the WebLINX dataset for real-world web navigation tasks. It uses a pixel-only approach (no HTML/DOM access) making it ideal for:

• Web automation tasks
• GUI interaction from screenshots
• Cross-platform browser control
• Accessibility applications

Files

File	Size	Description
encoder_model.onnx	88 MB	Vision encoder
decoder_model.onnx	183 MB	Text decoder
decoder_with_past.onnx	170 MB	Decoder with KV-cache
decoder_model_merged.onnx	728 MB	Merged decoder (for Transformers.js)

Usage

With ONNX Runtime (Python)

import onnxruntime as ort
from transformers import AutoProcessor
from PIL import Image

# Load processor
processor = AutoProcessor.from_pretrained("bollscoasts/pix2act-weblinx-base-onnx")

# Load ONNX model
session = ort.InferenceSession("encoder_model.onnx")

# Process image
image = Image.open("screenshot.png")
inputs = processor(images=image, return_tensors="np")

# Run inference
outputs = session.run(None, dict(inputs))

With Transformers.js (Browser)

import { pipeline } from "@xenova/transformers";

// Load the ONNX model
const pipe = await pipeline("image-to-text", "bollscoasts/pix2act-weblinx-base-onnx");

// Run inference on screenshot
const result = await pipe(image);
console.log(result);

Export Details

Exported using HotelBench ONNX export utilities.

The original model was missing the image processor config, which was restored from google/pix2struct-base to enable ONNX export.

Performance

• Browser (WebGPU): ~500-1000ms inference time
• Server (CUDA): ~100-200ms inference time
• Server (CPU): ~1-2s inference time

Citation

@article{pix2act,
  title={From Pixels to UI Actions: Learning to Follow Instructions via Graphical User Interfaces},
  author={Shaw, Peter and Joshi, Mandar and Cohan, Iz and Berant, Jonathan and Pasupat, Panupong and Shin, Hexiang and Gur, Izzeddin and Kwiatkowski, Tom and Zhang, Kenton Lee},
  journal={NeurIPS 2023},
  year={2023}
}

@article{weblinx,
  title={WebLINX: Real-World Website Navigation with Multi-Turn Dialogue},
  author={Lù, Xing Han and Kasner, Zdeněk and Reddy, Siva},
  journal={ICML 2024},
  year={2024}
}

License

Apache 2.0 - See original model for full license details.