beaunix/artwork-classifier-cnn

Melkov — Art Style Classifier

Part of the Aegis suite of specialized AI agents. This repository hosts the CNN vision model that powers Melkov (Art-Atelier), an agent that classifies the artistic style of a painting from an image. The agent's conversational/analytical layer runs separately on Qwen2.5-VL-7B, also hosted on Hugging Face.

Model

• Architecture: EfficientNetV2-Small (~80MB)
• Task: Art style classification (single-label image classification)
• Formats provided:
  • melkov-style-v2s.pt — PyTorch checkpoint. Use for fine-tuning or retraining.
  • melkov-style-v2s.onnx — ONNX export. Use for inference in any ONNX Runtime environment.

Training data

Trained on OpenBrush-75K by jaddai — 75,313 public domain artworks across 111 artists and 27 art movements, with structured VLM-generated metadata. Released under MIT.

Note: this repo was originally accessed via a re-upload/mirror (Trever896/openbrush-75k); the dataset's own citation credits jaddai as the original author, and that is the credit given here.

Usage

ONNX Runtime (inference only)

```python import onnxruntime as ort

session = ort.InferenceSession("melkov-style-v2s.onnx") outputs = session.run(None, {"input": input_tensor}) ```

PyTorch (inference or further training)

```python import torch

model = torch.load("melkov-style-v2s.pt") model.eval() ```

License

MIT. Training data (OpenBrush-75K) is also MIT-licensed and built entirely from verified public domain artworks, so no upstream art-copyright restrictions apply. Full credit to jaddai for compiling and captioning the dataset.

Related

• Project repo at: [GITHUB_REPO_URL](https://github.com/BeauBryanDev/melkov)
• Conversational layer (Qwen2.5-VL-7B): https://huggingface.co/spaces/beaunix/melkov