# LoRA Dataset Studio V2

A self-hosted, browser-based workbench for building, curating, captioning and training LoRA and full image-model checkpoints, covering everything from raw image collection through fine-tuned model comparison in one interface.

LoRA Dataset Studio runs on the user's own machine and uses ai-toolkit for training and ComfyUI for local generation and checkpoint testing, giving a single browser interface for the whole lifecycle of a character, concept or style dataset. Its Image Bank reviews existing image folders in place, scores quality, groups duplicate or similar images, and lets a user search and build a shortlist to promote into a working dataset; curation tools handle cropping, mirroring, rotation, face-similarity checks, watermark masking and text detection.

Captioning can run through local vision models or JoyCaption, with configurable appearance policies and bulk or targeted re-captioning. Local training covers guided recipes for several model families, including SDXL, FLUX.1 and Qwen-Image, with queued jobs, checkpoint continuation and experimental slider LoRAs, and a review stage generates and ranks checkpoints across seeded prompt batches. The core application and its public plugins are free to use under a noncommercial license, with training on a rented GPU and some external APIs incurring their own separate charges.

Website: https://github.com/perfectgf/lora-dataset-studio