lora-caption-dataset

What problem does it solve?

LoRA training datasets often fail reproducibility when captions are missing or inconsistent, because downstream training depends on a stable trigger-token format across every image.

Core Features & Use Cases

• Caption generation or validation: Creates or checks the <image-basename>.txt caption file for every training image, optionally overwriting as requested.
• Trigger-token enforcement: Guarantees each caption includes the required trigger token in the correct, deterministic position for the training pipeline.
• Class-consistent formatting: Applies character/product/style/pose/environment templates so captions follow uniform structure for downstream training.
• Noise reduction: Strips common caption leakage like model names, watermarks, and generic phrases to keep captions clean and usable.

Quick Start

Use the skill to generate standardized captions for all images in your LoRA dataset folder, ensuring every caption includes the derived trigger token and matches the class template you specify.