Albumentations

Published a new technical page: AlbumentationsX vs Albumentations MIT: Practical Benefits This page explains what AlbumentationsX adds in practice. albumentations.ai/docs/albumenta…

New @albumentations docs guide: how to choose augmentations. Core idea: each transform = an invariance assumption — then step-by-step pipeline design, tuning, diagnostics, when aug hurts. albumentations.ai/docs/3-basic-u…

Now in UI tool at explore.albumentations.ai you can see exact parameters that were used to generate each transformed image.

AlbumentationsX 2.1.1: github.com/albumentations… lightweight exact params — save_applied_params → serializable (name, config) per step + from_applied_transforms replay; LetterBox; Speedup in Normalize per image, RandomBrightnessContrast,

Albumentations 2.1.0 🚀 🆕 Custom apply targets: add apply_to_<key> (labels, heatmaps…) github.com/albumentations… 🆕 8 new transforms: - AtmosphericFog - ChannelSwap - FilmGrain - Halftone - LensFlare - Vignetting - GridMask - WaterRefraction

Not "flip, rotate, color jitter" — the real thing. Vladimir breaks down a decade of practical lessons: two augmentation regimes (in-distribution vs regularization), the one non-negotiable rule (label preservation), silent label corruption in detection/pose tasks, and how to match policy strength to model capacity. { author: @viglovikov } dev.to/viglovikov/ima…

@iglovikov/what-is-image-augmentation-4d31dcb3e1cc" target="_blank" rel="nofollow noopener">medium.com/@iglovikov/wha…

Albumentations 2.0.20 🚀 github.com/albumentations… - Perspective up to 2.7× faster on video batches - HueSaturationValue 1.2× faster. New user_data target: pass - camera intrinsics, - captions, LiDAR through your pipeline. pip install albumentationsx==2.0.20

New subreddit for Albumentations 🆕 Reddit → reddit.com/r/Albumentatio… GitHub, Discord, LinkedIn, X, email — now Reddit too.

🚀 Albumentations 2.0.19 is out github.com/albumentations… Two improvements for real inference pipelines ⚡ Less memory copying -> faster transforms 🔁 Deterministic TTA Explicit control over symmetry elements in D4 and flips Built-in .inverse() for clean prediction restoration.

AlbumentationsX 2.0.18 🚀 github.com/albumentations… PhotoMetricDistort + up to 40× faster multi-channel transforms (OpenCV 4.13). Satellite & medical folks—update.

Albumentations 2.0.17 is out. github.com/albumentations… Main change - added support for oriented bounded boxes to most of the geometric transforms. Also added `cxcywh` box format <- unnormalized YOLO.

Tweet 7/7 🧡 Sustainability note: AlbumentationsX is dual-licensed: • AGPL — open source & research • Commercial license — proprietary products Commercial use 👉 albumentations.ai/pricing/ Personal use? A coffee-sized sponsor helps a lot ☕ 👉 github.com/sponsors/album… Thanks 🙏

Tweet 6/7 This change closes an entire class of long-standing issues. It trades a bit of automation for: • Deterministic installs • Fewer broken environments • Fewer “why did this suddenly break?” moments Infrastructure should be boring. This helps.

🧵 AlbumentationsX 2.0.14 is out — and it fixes a long-standing OpenCV headache github.com/albumentations… For years, we tried to be “smart” about OpenCV dependencies. This release intentionally does less — and that’s why it’s better 👇