Training Captioner

Rulz Training Captioner

A Windows desktop app for captioning image datasets (LoRA / fine-tune training data) with a local vision LLM. Nothing is uploaded anywhere: the model runs on your own GPU(s).

Caption grid, dark theme

What it does

Load images or whole folders, shown as a thumbnail grid with an editable caption box under every image.
Captions are stored as .txt files next to each image (photo.png → photo.txt), the format trainers like kohya_ss, OneTrainer, ai-toolkit and SimpleTuner expect.
Edit any caption by hand; the .txt file is created or updated when you click away from the box.
Caption everything (or a selection) with a vision model through llama.cpp, or any OpenAI-compatible server.
Save your own system prompts for different caption styles.
Multi-GPU: split one big model across cards, or run one copy per card and caption images in parallel.
Dark theme (default) and light theme.

Start here

If you want to…	Read
Install and run the app	Installation
Caption your first folder	Quick Start
Learn the grid, editing and shortcuts	Using the Grid
Write and save caption prompts	Prompts
Pick a model and set up llama.cpp	Models and Backends
Use two or more GPUs	Multi-GPU
Look up every setting	Settings Reference
Fix a problem	Troubleshooting

Requirements

Windows 10/11 and Python 3.10+ (used by run.bat)
An NVIDIA GPU is recommended (CUDA builds of llama.cpp are what the app is tested with)
A vision-capable GGUF model and its mmproj file, or a running OpenAI-compatible server

Recent versions

Version	What changed
1.2.0	Faster captioning (pictures are shrunk while they load and the next one is prepared while the model works); busy-server errors are retried; a caption you type during Caption All is no longer overwritten; Start Servers with the same settings doesn't reload the model.