Character Extraction & AI Tagging
Extract sprites and auto-generate Danbooru-compatible tags for LoRA and LyCORIS training.
Character Extraction & AI Tagging is the professional-grade tool for visual novel developers and machine learning engineers who need clean, labeled character sprites. Upload any game screenshot, manga panel, or anime frame and our pipeline automatically detects character regions using Tachie clustering, separates them from backgrounds, and generates a structured tag list fully compatible with Booru-style datasets. The output is ready to train Stable Diffusion LoRA and LyCORIS models using tools like kohya-ss, EveryDream2, and SimpleTuner — no manual annotation required.
How Character Extraction & Tagging Works
Our pipeline chains four ONNX-powered modules that run entirely in your browser. No server upload, no account, no waiting queue — just drop your image and get a labeled dataset in seconds.
- Upload a sprite sheet, game screenshot, or illustration (JPG, PNG, WebP, up to 20 MB) 2. The Tachie clustering engine segments pixel regions and draws bounding boxes around each detected character 3. Each character is isolated and cropped into an individual transparent PNG with clean edges 4. Every crop is passed through a WD14-compatible tagger that generates quality, style, and attribute tags 5. Download a ZIP containing character PNGs and their .txt sidecar files — ready for LoRA training
✨ Key Features
- ✓ Tachie sprite clustering — detects multiple characters per frame
- ✓ Danbooru-compatible tag output (quality, artist-style, character attributes)
- ✓ Lineart & edge detection for clean sprite borders
- ✓ Truncated image detection — flags incomplete sprites automatically
- ✓ ONNX-powered, runs 100% in-browser with no server upload
- ✓ Export: PNG crops + TXT/JSON tags compatible with kohya-ss & EveryDream2
🎯 Built For
- → Building LoRA training datasets from visual novel and game assets
- → Automating sprite tagging for large art libraries
- → Creating labeled datasets for anime-style character classifiers
- → Extracting characters from manga panels for style-transfer pipelines
- → Quality-control workflows for AI illustration studios
- → Rapid prototyping of character recognition and classification systems
Frequently Asked Questions
What tag format does the output use? ▼
Tags are output in Danbooru-style flat format — one tag per line — fully compatible with kohya-ss, EveryDream2, and SimpleTuner. Each detected character gets its own .txt sidecar file alongside its cropped PNG.
How does it handle multiple characters in one image? ▼
The Tachie clustering algorithm identifies distinct character regions using color segmentation and bounding-box detection. Each region is extracted as a separate PNG with its own tag file — ideal for sprite sheets with multiple characters.
Is my artwork uploaded to a server? ▼
No. All inference runs locally via ONNX Runtime Web (WebAssembly) inside your browser. Your artwork never leaves your device — no accounts, no uploads, no data retention.