Winnow
Winnow audits a batch of training photos for duplicates, train/test leaks, blur, watermarks, AI-generated images and label mistakes, then has Claude decide which ones to remove.
Your photos are never shown publicly and are never used to train anything. Please don't upload anything private.
Kept for 30 days, in this browser only.
What Winnow checks for
Upload photos above, or load the example set, then press Scan to see your results here.
Test photos that also appear in your training photos. The model has already seen them, so its test score will look better than it really is.
Fix: delete the training copy of each pair (the download above does this), then re-measure your test score.
Train/Test Leaks — none found.
Your test photos look like a different kind of picture than your training photos, so a good test score may not carry over to real use.
Fix: add test photos that look like what the model will really see, or add training photos like the odd ones.
Train/Test Shift — none found.
Two files that are really the same picture, keeping both can trick a model into looking more accurate than it really is.
Fix: keep one photo from each pair. The download above drops the extras Claude marked Remove.
Duplicate Photos — none found.
Photos taken seconds apart on one camera. Their pixels can differ while the moment is the same, so spreading them across train and test inflates the score.
Fix: keep one photo per moment, and never split a moment between train and test.
Same-Moment Shots — none found.
Too out of focus to be useful for training.
Fix: drop them, or retake them if that class is short on examples.
Blurry Photos — none found.
Photos with a hidden "rotate me" tag. If your labels were drawn on the unrotated photo, they may no longer line up. Add your label files above and Winnow checks this for you; otherwise these are worth checking by hand.
Fix: open a few in your labeling tool. If the boxes are off, re-save the photos with rotation applied (for example Pillow's ImageOps.exif_transpose) and redraw those boxes.
Rotation Problems — none found.
The same logo, watermark or frame in many photos (a model can learn it instead of the subject), or a stock agency named in a photo's own metadata (its license may not allow training).
Fix: crop the overlay out or replace those photos, and check the license of any stock photo before you train on it.
Watermarks & Stock Photos — none found.
Photos whose own metadata says an AI tool made them. Winnow only reads what a file declares, so a photo with no marker isn't proof it's real.
Fix: decide whether generated images belong in your training set, and keep them out of your test set.
AI-Generated Photos — none found.
Boxes that leave the photo, have no area or repeat, label sizes that don't match the photo, and photos with no label.
Fix: open the flagged photos in your labeling tool and fix those boxes.
Label Problems — none found.
Files that failed to open, likely corrupted or not actually images.
Fix: re-export or re-download these files, a model can't learn from them.
Couldn't Be Scanned — none found.