LEGATO

LEGATO: Large-Scale End-to-End Generalizable Approach to Typeset OMR. This model performs Optical Music Recognition (OMR) on typeset sheet music images, transcribing them directly to ABC notation.

🔗 Try it: Interactive Demo | Leaderboard

⚠️ Important: This model must be used with the legato codebase. It cannot be loaded with standard Transformers pipelines alone due to the custom LegatoModel architecture.

Model Details

• Developed by: Guang Yang, Victoria Ebert, Nazif Tamer, Brian Siyuan Zheng, Luiza Pozzobon, Noah A. Smith
• Model type: Vision-language model for end-to-end OMR
• Architecture: Based on Llama 3.2 11B Vision (Mllama). Uses a frozen vision encoder and a trained text decoder that outputs ABC notation.
• License: MIT
• Paper: LEGATO: Large-Scale End-to-End Generalizable Approach to Typeset OMR (ICLR 2026)

How to Use

Installation

1. Clone the repository and install dependencies:

git clone https://github.com/guang-yng/legato.git
cd legato
pip install -r requirements.txt

Tested with Python 3.12 and CUDA 12.4.

2. Access requirements: This model loads the vision encoder from meta-llama/Llama-3.2-11B-Vision. Ensure you have accepted the Llama 3.2 license on Hugging Face.

Inference

import torch
from PIL import Image
from transformers import AutoProcessor, GenerationConfig
from legato.models import LegatoModel

# Load model and processor
model = LegatoModel.from_pretrained("guangyangmusic/legato")
processor = AutoProcessor.from_pretrained("guangyangmusic/legato")

# Move to GPU
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

# Load and process image
image = Image.open("path/to/sheet_music.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

# Generate ABC notation
generation_config = GenerationConfig(
    max_length=2048,
    num_beams=10,
    repetition_penalty=1.1
)

with torch.no_grad():
    outputs = model.generate(**inputs, generation_config=generation_config)

# Decode output
abc_notation = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(abc_notation)

Half-Precision Inference (Reduced Memory)

model = LegatoModel.from_pretrained("guangyangmusic/legato")
model = model.to("cuda").half()  # Use FP16

Batch Inference

images = [Image.open(p).convert("RGB") for p in image_paths]
inputs = processor(images=images, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    outputs = model.generate(**inputs, generation_config=generation_config)

abc_outputs = processor.batch_decode(outputs, skip_special_tokens=True)

Output

The model outputs ABC notation transcriptions. ABC can be converted to MusicXML using the conversion utilities in the legato codebase.

Intended Use

• Primary use: Transcribing typeset sheet music images (piano scores, orchestral parts, etc.) to ABC notation
• Audience: Researchers and practitioners in music information retrieval, digital humanities, and music technology
• Out-of-scope: Handwritten notation, audio-to-score transcription, real-time inference without GPU

Limitations

• Trained primarily on synthetic typeset data; performance may degrade on handwritten scores, low-quality scans, or unusual layouts
• Requires significant GPU memory (~20GB+ for full precision; use --fp16 for lower memory)
• Depends on access to meta-llama/Llama-3.2-11B-Vision for the vision encoder
• Maximum generation length: 2048 tokens (default)

Training

Trained on PDMX-Synth with DeepSpeed ZeRO-2. For training and validation instructions, see the legato repository.

Evaluation

The codebase provides evaluation scripts for:

• TEDn – Tree Edit Distance on MusicXML
• OMR-NED – Normalized Edit Distance via musicdiff

See the README for evaluation commands.

Citation

@inproceedings{yang2026legato,
  title={{LEGATO}: Large-scale End-to-end Generalizable Approach to Typeset {OMR}},
  author={Guang Yang and Victoria Ebert and Nazif Can Tamer and Brian Siyuan Zheng and Luiza Amador Pozzobon and Noah A. Smith},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  eprint={2506.19065},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2506.19065}
}

Related Models

Recommended: Use legato (this model). The small variant is mainly for baseline comparisons and is less efficient.