LoraTag - AI-Powered Image Captioning for LoRA Training | Batch Dataset Preparation Tool

What

LoraTag is a web-based image captioning tool for preparing LoRA training datasets. It uses GPT-4 Vision or OpenAI Vision, depending on the page section, to generate natural-language captions for images and export matching .txt files used by LoRA training workflows.

It is aimed at AI artists, LoRA creators, and fine-tuning researchers working with Stable Diffusion, SDXL, SD3, and FLUX-related training pipelines. Its core workflow is upload images, choose caption detail settings, generate captions in batch, review or edit them, and download training-ready files; positioning appears to be a browser-first dataset preparation tool focused on faster and more consistent captioning than manual methods or simple taggers.

Features

• Batch image captioning — Processes many images at once to reduce the manual effort of writing captions one by one.
• Multiple caption detail levels — Lets users choose shorter or more descriptive outputs depending on training goals and token constraints.
• Natural-language caption generation — Produces descriptions that include subjects, composition, style, lighting, color, and context rather than only comma-separated tags.
• Training-ready .txt export — Outputs one caption file per image in a format intended for common LoRA trainers such as kohya_ss, EveryDream, SimpleTuner, ai-toolkit, and related tools.
• In-browser review and editing — Allows users to inspect, refine, or regenerate captions before downloading the dataset.
• Folder and dataset organization support — Supports folders and preserves directory structure, which helps maintain larger training datasets during export.

Helpful Tips

• Check plan and quota details carefully — The source content shows inconsistent limits and pricing across sections, so buyers should verify current image allowances, quality levels, and plan features on the live product before committing.
• Use cleaned and consistently framed images — Caption quality may be more useful for training when the dataset already has consistent cropping, subject focus, and minimal noise.
• Match caption length to the training objective — Shorter captions may suit trigger-word workflows, while standard or detailed captions are more appropriate when composition, style, and relationships matter.
• Review captions before export — Even with strong vision models, manual review is still important for edge cases, naming consistency, and token budget control.
• Confirm data handling requirements — The page states images are processed in memory, sent to OpenAI’s Vision API, and then deleted, so teams with strict privacy requirements should validate that workflow against internal policies.