Target category
Controls which caption variant becomes the final_caption in the output. The Caption tool exposes the full set of categories; these are the ones you reach for most:
- identity — for a person/character LoRA; keep this as the default when the face is the concept.
- wardrobe — for clothing/outfit LoRAs where the garment is the focus.
- setting — for scene LoRAs that reward wide, establishing framing.
The category also shapes what good input looks like: an identity set wants sharp, varied face angles; a setting set wants wide, high-resolution scenes. Curate scores images with a related category taxonomy, so picking the same concept in both steps keeps captioning and selection aligned.
Pick a category, see what to feed it
The category you train for changes both the caption and the kind of images that make a strong dataset. This widget is driven by the very same scoring guidance the Curate tool uses, so it reflects real behaviour:
- What it trains
- Person / character LoRA — strict blur floor, rewards a single centred face.
- What good input looks like
- The face is the concept: favour clear, sharp face shots at varied angles (head-on, 3/4, profile) and expressions, plus a few half/full-body frames where the face is still visible. Faceless crops don't teach identity — enable Face Clustering, then filter with “Require a known face”.
- Suggested set size
- ~15–30 sharp, varied images
Where it shows up
The concept appears in two places. In the Caption tool, Target Category
selects which variant is written as the final_caption. In Curate the
same idea biases the image scorer toward the framing each concept needs — that's
the behaviour the widget above previews. The two tools share the core categories
(identity, wardrobe, setting); Curate concentrates on the ones that most change
what a good dataset looks like.
Try it
- Open the Caption tool and expand the Target Category info panel — it reads from the same source as this page.
- Head to Curate to score a folder against your chosen category.