Normalizing Flows

Normalizing flows are generative AI models that learn an invertible mapping between a simple probability distribution, usually Gaussian noise, and a complex data distribution. Unlike many generative models, they can both create samples and calculate exact data likelihoods efficiently. This makes them useful for density estimation, uncertainty modeling, and learning structured latent spaces, as explained in the foundational normalizing flows review. (arxiv.org)

Normalizing Flowsの仕組み#

フローは、一連の可逆的なニューラルネットワーク変換を適用します。

1. 単純な基底分布からポイントをサンプリングする。
2. 複数の可逆層を通じてそれを変換する。
3. ヤコビアン行列式を使用して、各層がどのように確率密度を拡大または縮小するかを追跡する。
4. 観測データの確率を計算する際に、変換を逆転させる。

The Pyro normalizing flow tutorial provides practical examples of sampling and density evaluation. Although the name sounds similar, normalizing flows are not the same as feature normalization or batch normalization. Here, “normalizing” means transforming a complex distribution into a standard one.

Traditional designs require carefully structured invertible layers. Recent Free-form Flows research relaxes this restriction, while a 2024 universality analysis of coupling-based flows explains why affine coupling layers remain effective. (proceedings.mlr.press)

実社会での応用#

• Industrial Anomaly Detection: A flow can model embeddings from defect-free products and flag unlikely samples during visual inspection. However, likelihood alone is not always a reliable out-of-distribution score, as shown by NeurIPS research on flow likelihood failures. Validate results with task-specific metrics and representative abnormal data. (proceedings.neurips.cc)
• Medical Imaging: The 2024 transcranial ultrasound flow model uses normalizing flows for faster reconstruction and uncertainty estimation. Similar techniques can support medical image analysis where predictions need confidence ranges. (proceedings.mlr.press)
• Synthetic Data and Calibration: Flows can generate structured synthetic data or model prediction errors. The 2024 work on normalizing flows for conformal regression demonstrates more adaptive uncertainty intervals. Generated vision data can be annotated, trained, and deployed through the Ultralytics Platform. (proceedings.mlr.press)

Normalizing Flowsと関連手法の比較#

Flow matching usually trains a continuous velocity field with a regression objective, as detailed in Meta’s Flow Matching guide. Traditional normalizing flows instead emphasize invertible transformations and direct likelihood optimization. Rectified flow seeks straighter transport paths, while diffusion models generate data through iterative denoising. Normalizing flows also differ from Generative Flow Networks, which learn policies for constructing discrete objects, and from GANs, which do not normally provide exact likelihoods. (ai.meta.com)

最近の動向とベストプラクティス#

Transformer-based architectures have renewed interest in flows. The 2025 TarFlow study reported image generation competitive with diffusion approaches, while Jet modernized coupling flows with Vision Transformers. In 2026, regression-based flow training connected normalizing flows with flow-matching-style objectives, and SESaMo incorporated exact physical symmetries. (proceedings.mlr.press)

ビジョンアプリケーションにおいては、生のピクセルではなく高レベルの埋め込みをモデル化するのが実用的なアプローチです。

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
embeddings = model.embed("https://ultralytics.com/images/bus.jpg")
print(embeddings[0].shape)

These Ultralytics YOLO26 embeddings can become inputs to a separately trained flow for density estimation or anomaly scoring. Use careful data preprocessing and evaluate likelihood alongside downstream performance rather than treating it as a complete quality measure.