Ayase
Modular media quality metrics for video, image, and audio datasets.
Work in progress - APIs and module interfaces may change before 1.0.
What It Does
Ayase runs quality assessment modules over a dataset and writes structured per-sample metrics. 375 modules produce 511 metrics across 21 categories (NR-IQA, FR-IQA, NR-VQA, temporal, motion, audio, face, safety, aesthetics, text-video alignment, and more). Modules are independent - pick only what you need.
Full metric catalog: METRICS.md. Pretrained model catalog: MODELS.md.
Install
pip install ayase
Ayase is distributed as a single install. Runtime dependencies are managed by the project itself, and model weights are downloaded and cached on first use.
Isolated installation
Applications that already have their own PyTorch, NumPy, or media dependency constraints can keep Ayase completely outside their environment:
pip install ayase --no-deps
from ayase_client import AyasePipeline
pipeline = AyasePipeline(modules=["basic", "metadata", "motion"])
results = pipeline.run("./my_dataset")
pipeline.export("report.json")
pipeline.close()
This installs the same Ayase wheel without ML dependencies. On first use the
client creates a private full-runtime venv and starts a loopback worker. Media is
passed by local path. Regular pip install ayase remains unchanged.
Optional check: ayase-client doctor. To use an existing runtime, set
AYASE_RUNTIME_PYTHON.
CLI
ayase scan ./dataset # default balanced pipeline
ayase scan ./dataset --deep # run every discovered module
ayase scan ./dataset --modules metadata,basic_quality # selected modules
ayase help # list every metric and provider
ayase help rqvqa_score # metric/module models, config, and usage
ayase modules list # show all 375 modules
ayase modules check # import/dependency readiness
ayase filter ./dataset --min-score 70 --output ./good # filter by quality
ayase stats ./dataset # dataset statistics for images/video
ayase tui # terminal UI
Python API
from ayase import AyasePipeline
pipeline = AyasePipeline(modules=["basic", "metadata", "motion"])
results = pipeline.run("./my_dataset")
for path, sample in results.items():
qm = sample.quality_metrics
if qm:
print(f"{sample.path.name}: technical={qm.technical_score} blur={qm.blur_score}")
pipeline.export("report.json") # also: report.csv, report.html
Configuration
ayase.toml in project root:
[general]
parallel_jobs = 8 # concurrency hint passed to capable modules/backends
[pipeline]
modules = ["metadata", "basic_quality", "motion"]
[output]
default_format = "json"
artifacts_dir = "reports"
Custom Modules
from ayase.models import QualityMetrics, Sample, ValidationIssue, ValidationSeverity
from ayase.pipeline import PipelineModule
import cv2
class BlurCheck(PipelineModule):
name = "blur_check"
description = "Flag blurry frames via Laplacian variance"
default_config = {"threshold": 100.0}
def process(self, sample: Sample) -> Sample:
img = cv2.imread(str(sample.path), cv2.IMREAD_GRAYSCALE)
if img is None:
return sample
score = float(cv2.Laplacian(img, cv2.CV_64F).var())
if sample.quality_metrics is None:
sample.quality_metrics = QualityMetrics()
sample.quality_metrics.blur_score = score
if score < self.config.get("threshold", 100.0):
sample.validation_issues.append(
ValidationIssue(
severity=ValidationSeverity.WARNING,
message=f"Blurry ({score:.0f})",
)
)
return sample
Modules auto-register via __init_subclass__. Config is available as self.config.
Development
git clone <repo-url> && cd ayase
pip install -e ".[dev]"
pytest # 8000+ tests, ~4 min
pytest tests/ --full # with ML model loading
License
Ayase's own code is MIT. Model weights downloaded at runtime carry their own licenses - see MODELS.md.
Some metrics run research code vendored under ayase/vendor (see
the inventory), and four of those components are not
permissive. Running one of these metrics places its component's licence on your use
of the result, whatever Ayase's own licence says. Each affected module declares what
it runs and logs a notice at setup:
| Metric | Vendored component | Licence |
|---|---|---|
chronomagic, dynamics_controllability, physics, video_edit_motion_fidelity, vmbench_pas, vmbench_tcs |
CoTracker | CC BY-NC 4.0 - non-commercial only |
mj_video |
MJ-Video | no licence file upstream - no grant is stated |
vbench2 |
VBench 2.0 with its vendored YOLO-World and CoTracker | Apache-2.0, plus GPL-3.0 and CC BY-NC 4.0 inside |
Every other vendored component is MIT, BSD or Apache-2.0 and imposes nothing beyond attribution, which the retained licence files provide.
This is a disclosure, not a resolution. The plan for 1.0 is to replace these components with implementations Ayase can license itself, so that the whole distribution is MIT in substance and not only in name.
1 maintainer
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file ayase-0.1.75.tar.gz.
File metadata
- Download URL: ayase-0.1.75.tar.gz
- Upload date:
- Size: 13.3 MB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.10.11
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
ae4b420abc25ca3ec19d98f70bd966d49e2e9f99876f05f8d523072073a0169d
|
|
| MD5 |
588df170a0e67d734a733936e556ad2e
|
|
| BLAKE2b-256 |
c18928046bbbb64a32748a1499c458d9f50d7b71d2bdb6247d91bb72ee964aa7
|
File details
Details for the file ayase-0.1.75-py3-none-any.whl.
File metadata
- Download URL: ayase-0.1.75-py3-none-any.whl
- Upload date:
- Size: 15.7 MB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.10.11
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
ca5f4691378731bec2dff8892c88293277530fc2555a498239754a249b968c55
|
|
| MD5 |
74e4c734b6010fcb48f54bade7235b5f
|
|
| BLAKE2b-256 |
59e020b567b8f2a7f41cca8fc2ad5bbd26c8be3a2140bb144d4b10ac4c1b365c
|
