// the find
cleanlab/cleanvision
Automatically find issues in image datasets and practice data-centric computer vision.
CleanVision scans raw image datasets for quality problems — duplicates, blur, exposure issues, bad aspect ratios — before you train anything. It's a pre-training audit tool, not a model or a label checker; for label issues you need the sibling cleanlab package. Aimed at CV practitioners who are tired of discovering garbage data after a long training run.
The issue detection runs on raw pixels without requiring a trained model, so you can use it at dataset assembly time rather than after the fact. Near-duplicate detection via perceptual hashing is genuinely useful and catches resized or slightly cropped copies that exact-match misses. The dataset abstraction layer supports local files, HuggingFace datasets, and Torchvision datasets through a single API, which means it slots into whatever pipeline you're already using. The issue manager architecture is pluggable — you can register custom issue types without forking the library.
Nine issue types is a short list; there's no detection for things like mislabeled content, out-of-distribution images, or NSFW material that matter a lot in real dataset audits. Performance on large datasets is undocumented — no benchmarks for how it behaves on a million images, which is where dataset quality problems tend to live. Last commit was January 2026 and the repo has low fork activity, which suggests Cleanlab's commercial product is where active development actually happens. The scoring system produces numerical quality scores but gives you no guidance on what threshold to actually use for filtering, leaving that call entirely to the user.