cleanvision

Why we included this project

Before training a computer vision model, it helps to know what is actually in your image folder, and CleanVision does that check for you. Point it at any directory of image files and it scans the raw pixels for the usual failure modes: blurry shots, under- or over-exposed images, near and exact duplicates, low-information frames, grayscale stragglers, and odd aspect ratios or sizes. The API is small, an Imagelab object with find_issues() and report() calls, and you can restrict the audit to just the issue types you care about. It accepts common image formats and handles local folders, torchvision datasets, or Hugging Face pulls the same way, so it works as an early sanity check inside an existing pipeline. If your problem is mislabeled images rather than bad pixels, the sibling cleanlab package is the better tool; for cleaning the images themselves, this is a fast pre-processing step worth running first.