README.md

---
tags:
- onnx
- transformers.js
- image-classification
- creativity-assessment
- beit
license: apache-2.0
datasets:
- figural-drawings
metrics:
- accuracy
pipeline_tag: image-classification
---

# OCSAI-D Web (ONNX Quantized)

This is a quantized ONNX version of the [POrg/ocsai-d-web](https://huggingface.co/POrg/ocsai-d-web) model, optimized for web deployment with Transformers.js.

## Model Description

This model assesses originality/creativity in figural drawings. It's a fine-tuned BEiT-large model that outputs a regression score indicating the creativity level of the input drawing.

## Model Details

- **Base Model**: microsoft/beit-large-patch16-224-pt22k-ft22k
- **Task**: Image regression (creativity scoring)
- **Input**: 224x224 RGB images
- **Output**: Single regression score (0-1 range)
- **Quantization**: INT8 dynamic quantization
- **File Size**: ~300MB (vs 1.1GB original)

## Usage with Transformers.js

```javascript
import { pipeline } from '@xenova/transformers';

// Load the model
const classifier = await pipeline('image-classification', 'your-username/ocsai-d-web-onnx');

// Run inference on an image
const result = await classifier('path/to/drawing.jpg');
console.log(result);
```

## Important Notes

This model is specifically designed for creativity assessment of figural drawings. The output is a single regression score that needs to be post-processed according to the original paper's methodology.

## Original Model

Based on [POrg/ocsai-d-web](https://huggingface.co/POrg/ocsai-d-web) - please refer to the original model for citation information and detailed usage instructions.

## Performance

The quantized model provides significant size reduction (~4x smaller) while maintaining compatibility with Transformers.js for browser-based inference.