Patch-ioner_talk2dino_decap_COCO_Captions - Patch-ioner Configuration

This repository contains a pre-trained DECAP model from the Patch-ioner framework for dense image captioning and controllable visual description.

📝 Paper Information

Title: "One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework"
Authors: Lorenzo Bianchi, Giacomo Pacini, Fabio Carrara, Nicola Messina, Giuseppe Amato, Fabrizio Falchi
ArXiv: https://arxiv.org/abs/2510.02898 Project Page: https://paciosoft.com/Patch-ioner/ Code: https://github.com/Ruggero1912/Patch-ioner

🎯 Model Overview

• Model Type: DECAP
• Configuration: mlp.karpathy.yaml
• Vision Backbone: dinov2_vitb14_reg
• Language Model: GPT-2
• Input Resolution: 518x518
• Prefix Size: 768

DeCap Configuration

• Memory Bank Size: 591,753 entries
• Projection Type: /raid/datasets/im2txtmemories/coco_train_karpathy.json
• Linear Talk2DINO: False

📊 Performance

| Task | METEOR | CIDEr | SPICE | |------|--------|-------|-------|
| Image Captioning | 0.239 | 0.885 | 0.182 |
| Narratives | 10.700 | 27.900 | 12.600 |

📈 Detailed Results

Image Captioning Results

• METEOR: 0.2393
• CIDEr: 0.8846
• SPICE: 0.1821
• BLEU_4: 0.2364
• ROUGE_L: 0.4854
• CLIP-S: 0.7602

Narratives Results

• METEOR: 10.7000
• CIDEr: 27.9000
• SPICE: 12.6000
• BLEU_4: 2.5000
• ROUGE_L: 23.2000
• CLIP-S: 68.0000

🚀 Quick Start

from patchioner import Patchioner
from transformers import AutoModel

MODEL_ID = "Ruggero1912/Patch-ioner_talk2dino_decap_COCO_Captions"

# Option 1: Using Patchioner's native API
model_native = Patchioner.from_config(MODEL_ID)
print("Model loaded via native Patchioner API.")

# Option 2: Using Hugging Face Transformers AutoModel
model_hf = AutoModel.from_pretrained(MODEL_ID, trust_remote_code=True)
print("Model loaded via Hugging Face Transformers AutoModel.")

# For full usage examples, including inference, please refer to the model's
# GitHub repository: https://github.com/Ruggero1912/Patch-ioner

📁 Repository Contents

• config.yaml: Model configuration file

• coco_karpathy-009.pt: Pre-trained model weights

• coco_captions_text_embeddings-B16-ViT-B.16-591753.h5: Memory bank for text projection- README.md: This file

🔧 Installation

pip install git+https://github.com/Ruggero1912/Patch-ioner

💡 Usage Examples

Refer to the Patch-ioner GitHub repository for updated usage examples.

🎛️ Model Configuration

• Prefix Size: 768
• Memory Bank Size: 591,753 entries
• Normalization: True

📈 Training Details

• Training Dataset: COCO Captions
• Training Epochs: TBD
• Batch Size: TBD
• Learning Rate: TBD
• Optimizer: AdamW

📚 Citation

If you use this model in your research, please cite our paper:

@misc{bianchi2025patchcaptionallunified,
      title={One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework},
      author={Lorenzo Bianchi and Giacomo Pacini and Fabio Carrara and Nicola Messina and Giuseppe Amato and Fabrizio Falchi},
      year={2025},
      eprint={2510.02898},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.02898},
}

🤝 Contributing

We welcome contributions to improve the Patch-ioner framework. Please see the main repository for contribution guidelines.

📄 License

See the main repository for detailed license information.

🐛 Issues and Support

For issues related to this model or the Patch-ioner framework, please:

1. Check the main repository for existing issues
2. Open a new issue with detailed information about your problem
3. Contact the authors.

🔗 Related Models

Explore other Patch-ioner model configurations:

• Patch-ioner_mlp - MLP-based DeCap model
• Patch-ioner_viecap - VieCap controllable captioning
• Patch-ioner_clipcap - ClipCap integration

More models available in Ruggero1912's models

---

This model is part of the Patch-ioner framework for dense image captioning and controllable visual description.