When a search returns hundreds of images, sorting them by hand is the slow part. MultiScraper's optional AI Filter does the first pass for you: it scores every image and keeps only the ones that match a classifier you train on your own examples. It runs locally — nothing is uploaded.
Set it up
- Set up Python once. Double-click setup-python.bat. It installs the small local environment the filter runs in; no manual Python setup.
- Label a few examples. On the Label page, mark a handful of images as keep or drop. A few dozen of each is enough to start.
- Train the classifier. On the Train page, train a local model on your labels. It runs on your GPU, or on CPU if you have no GPU.
- Turn the filter on. Enable the AI Filter. From now on it scores every image as it downloads and keeps only the matches.
Why train your own
A generic tag is blunt: it cannot tell a clean reference image from a cluttered one, or your target style from a near miss. Because you teach the filter with your own keep/drop examples, it learns your standard. That makes it especially useful when building a LoRA dataset, where consistency matters.
Using it day to day
- Start with a wide tag search so the filter has plenty to score.
- Let it keep the matches, then do a quick manual pass in the grid for edge cases.
- Re-train occasionally with new examples as your target sharpens.