A LoRA or fine-tune is only as good as its dataset. MultiScraper was built to make the collecting and cleaning part fast: search several art sites at once, keep only what fits, and export full-resolution originals with captions in a kohya-ready layout.
Build the dataset
- Collect candidates. Search a concept or style across pixiv, DeviantArt and kemono at once. Keep everything in the grid, then drop off-topic images with X.
- Let the AI Filter narrow it. Optional: turn on the AI Filter and train it on a few of your own examples so it keeps only images that match your concept.
- Export a clean folder. Save kept images to datasets\<name>. MultiScraper removes near-duplicates and keeps full-resolution originals.
- Add captions and a kohya layout. Choose caption files and the kohya folder layout on export, so the dataset is ready for a trainer with no reshuffling.
What you get on export
- Full-resolution originals — not thumbnails or re-encodes.
- Near-duplicates removed, so the trainer does not over-weight repeated images.
- Caption files next to each image when you enable them.
- kohya folder layout, ready to point a trainer at.
- Everything under
datasets\<name>, and still searchable offline by tag and source.
Good practice
Aim for a consistent subject or style with varied poses, framing and backgrounds. Review the captions before training — accurate tags matter more than volume. And collect responsibly: follow each site's terms of service and respect the rights of the artists whose work you train on.