The browser-based beast that turns chaotic image piles into perfectly tagged, ready-to-train datasets – faster than you can say "trigger word activated!"
Tired of wrestling with folders full of untagged images like a digital archaeologist? TagPilot swoops in like a supersonic jet, handling everything client-side so your precious data never leaves your machine (except when you politely ask an AI provider to peek for tagging magic). Private-ish, local-first, and zero server drama.
- Upload Shenanigans: Drag in single pics, or drop a whole ZIP bomb – it even pairs existing .txt tags like a pro matchmaker. Add more anytime; no commitment issues here.
- Trigger Word Superpower: Type your magic word once (e.g., "ohwx woman") and watch it glue itself as the VIP first tag on every image. Boom – consistent activation guaranteed.
- AI Tagging Turbo: Powered by Gemini 3.1 Flash Lite, Claude Sonnet 4.5, Grok 4.3, OpenAI GPT-5.4 Mini, DeepDanbooru, or WD1.4 – because why settle for one engine when you can have a fleet?
- Batch modes: Ignore (I'm good, thanks), Append (more tags pls), or Overwrite (out with the old!).
- Progress bar + emergency "Stop" button for when the API gets stage fright.
- Separate Tagging and Captioning Defaults: Set one default model for tags and another for captions. Tag with WD1.4, caption with Claude, or mix providers however your dataset pipeline demands.
- Proper Settings Cockpit: The top-left Settings panel now has dedicated Tagging Options, Captioning Options, and Crop Options sections for default models, output limits, editable system prompts, crop ratio, and crop width.
- Provider Key Table: Gemini, Grok, OpenAI, Claude, DeepDanbooru, and WD1.4 sit in one key/status table, so you can see which providers are ready instead of playing API-key hide-and-seek like it is 2009.
- Tag Viewer Cockpit: Collapsible dashboard showing every tag's popularity. Click the little × to yeet a bad tag from the entire dataset. Global cleanup has never felt so satisfying.
- Per-Image Playground: Clickable pills for tags, free-text captions, add/remove on the fly. Toggle between tag-mode and caption-mode like switching altitudes. Single-image Tag/Caption actions now show a tiny processing indicator, because silent buttons are rude.
- Crop & Conquer: Free-form cropping, preset LoRA-friendly ratios (1:1, 16:9, 3:2, 4:3, 21:9, 9:16, 2:3, 3:4), and a configurable output width. No more awkward compositions ruining your training.
- Duplicate Radar: 100% local hash detection – skips clones quietly, no false alarms from sneaky filename changes.
- Export Glory: One click → pristine ZIP with images + .txt files, ready for kohya_ss or your trainer of choice.
- Privacy First: Everything runs in your browser. API keys stay in your browser's local storage. Images are only sent to the AI provider you choose for tagging/captioning.
No servers, no npm drama – just pure single-file HTML bliss.
- Clone or download:
git clone https://github.com/vavo/TagPilot.git - Open
tagpilot.htmlin your browser. Done! 🚀
(Pro tip: For a fancy local server, run python -m http.server 8000 and hit localhost:8000.)
- Load Cargo: Upload images or ZIP – duplicates auto-skipped.
- Set Trigger: Your secret activation phrase goes here.
- Name Your Mission: Dataset prefix for clean exports.
- Configure Defaults: Use top-left Settings to choose default tag/caption models, prompts, limits, crop size, crop width, and provider keys.
- Tag/Caption All: Hit the button, choose how to handle existing text, and let the progress bar do its suspiciously useful job.
- Fine-Tune: Crop, manual edit, nuke bad tags globally.
- Deploy: Export ZIP and watch your LoRA soar.
- Vanilla JS + Tailwind (fast & beautiful)
- JSZip for ZIP wizardry
- Cropper.js for precision framing
- Web Crypto for local duplicate detection
- Multiple AI backends: Gemini, Claude, Grok, OpenAI, DeepDanbooru, and WD1.4
- Current cloud defaults:
gemini-3.1-flash-lite,claude-sonnet-4-5-20250929,grok-4.3, andgpt-5.4-mini - Tiny no-install Node test harness for provider payload sanity checks
- v2.0 refactors keep the single-file app intact while cleaning up provider calls, safe tag rendering, batch processing, object URL cleanup, stable image IDs, and shared tag utilities.
See CHANGELOG.md for the release notes.
Got ideas, bugs, or want to contribute? Open an issue or PR – let's make dataset prep ridiculously awesome together!
Happy training, pilots!