ayase 0.1.23

Modular media quality metrics toolkit

Ayase

Modular media quality metrics toolkit for video, image, and audio datasets.

Overview

• 324 modules, 354 quality metrics across visual, temporal, audio, perceptual, and safety categories
• Modular pipeline - modules compute raw values, downstream apps decide what to do with them
• CLI and Python API with profile-based configuration
• See METRICS.md for the full reference, MODELS.md for all pretrained weights

Installation

pip install ayase

Ayase ships with the shared runtime dependencies used by the bundled metrics. Some model implementations include source files under src/ayase/third_party/; model binaries are downloaded and cached automatically on first use when they are not packaged in the repository.

Quick Start

ayase scan ./my_dataset
ayase scan ./my_dataset --modules metadata,basic_quality,motion
ayase modules list
ayase modules check
ayase filter ./my_dataset --min-score 70 --output ./filtered

from ayase import AyasePipeline

ayase = AyasePipeline(modules=["basic", "aesthetic", "motion"])
results = ayase.run("./my_dataset")

for path, sample in results.items():
    if sample.quality_metrics:
        print(f"{sample.path.name}: score={sample.quality_metrics.technical_score}")

ayase.export("report.json")

Configuration

Create ayase.toml in your project root:

[general]
parallel_jobs = 8

[pipeline]
modules = ["metadata", "basic_quality", "motion"]

[output]
default_format = "json"
artifacts_dir = "reports"

Writing Plugins

from ayase.models import Sample, ValidationIssue, ValidationSeverity
from ayase.pipeline import PipelineModule

class MyCheck(PipelineModule):
    name = "my_check"
    description = "Custom quality check"
    default_config = {"threshold": 0.5}

    def process(self, sample: Sample) -> Sample:
        # Your logic here
        return sample

ayase scan ./data --modules metadata,my_check

Development

git clone <repo-url> && cd ayase
pip install -e ".[dev]"
pytest

License

MIT - see LICENSE.

Model licenses: ML models downloaded at runtime have their own licenses (Apache 2.0, BSD, CC-BY-NC, etc.). Users are responsible for compliance. See MODELS.md for the full catalog with license info. No model weights are bundled with this package.