NIN_Captioer — Local AI Captioning, Krea 2 LoRA Train & Test

NIN_Captioer is a Windows desktop tool for the full Caption → Train → Test loop. Start by preparing LoRA datasets with local Auto Caption and bilingual review, then train a native Krea 2 LoRA and verify results image.

Scope: Training is Krea 2 only (train on Krea-2-Raw).
Other bases (SD / SDXL / Flux, etc.) are not supported for training.
Tested on: Windows x64 + NVIDIA RTX 4090 only. Other GPUs / OS setups are untested.

Core Features

Auto Natural Language Captioning

• Seamlessly connects to LM Studio or Ollama via local APIs

• Uses local Vision-LLM models to generate detailed natural language captions suited for Flux / Krea2-style datasets

• Optional WD14 Danbooru tags (ONNX) for tag-style captioning

• Batch Auto Caption for images missing .txt, plus reCaption for the current image

• Caption Analysis: coverage, LoRA Health Score, category distributions

Real-Time Translation & Side-by-Side Comparison

• Instant side-by-side bilingual translation next to English captions

• Review, verify, and fine-tune AI captions without language barriers

• Edits on either side stay in sync (default target: Traditional Chinese)

Lora Train (Krea 2 only)

• Native Krea 2 LoRA trainer in-app (Raw train → Turbo apply / sample)

• Multi-job Start / Stop, progress log, loss chart, sample previews

• Loss spike markers and OOM detection surfaced in the UI and log

• Resource monitor (CPU / RAM / GPU)

• Low VRAM helpers: layer offload, quantization, gradient checkpointing

• Hugging Face model Download / Update, Python path probe and install helper

• LoRA export compatible with ComfyUI key format

Lora Test (ComfyUI)

• Install / start / stop ComfyUI from the app

• Generate with trained Krea 2 checkpoints, multi-prompt runs, and a result gallery

• Smoke-test LoRAs right after training without switching tools

Suggested workflow

1. Add your image folder → connect LM Studio / Ollama

2. Auto Caption → review with translation → Analyze Health Score

3. Lora Train (Krea 2 Raw / Turbo) → Start Train

4. Lora Test → start ComfyUI → pick checkpoint → generate & review gallery

Links

• GitHub Repository: BulbulLeung/NIN-Captioer