Related notes
A practical LoRA captioning guide
Separate the concept from changing details, with character, style and product examples.
Separate the concept from the scene
A caption describes the visible content of an image. For training, it can also distinguish the intended concept from changing details such as clothes, props, viewpoint or lighting. Do not assume the caption should describe only the part you want learned, or use one identical sentence blindly across the dataset.
Character captions describe variable details
Use a distinctive concept word where the trainer supports one, then describe the pose, clothing and surroundings that vary between images. Keep descriptions grounded in what is visible. A caption cannot compensate for a dataset that never shows the character in another setting or viewpoint.
Style captions should still identify the subject
When a dataset contains different subjects in one visual language, identify the subject and relevant setting rather than labeling every image only as a style. Review which textures, contours and palette choices define the intended treatment. Follow the selected trainer’s caption and trigger format.
Product captions distinguish object and context
Keep the concept consistent and describe the angle, surface and lighting around it. Describe changing props so they are not silently treated as permanent parts of the product. Avoid claiming invisible construction details or materials that are not supported by the image.
Review before exporting
The workspace currently supports manual caption editing. Automatic image descriptions are not connected. Open the caption editor after selecting images, review every entry and export the JSON draft. Keep the source images: the export contains metadata and text, not the image files or a trained adapter.