Target style
Determines how tags are classified and budgeted for the caption.
- photo optimises for realism models using natural-language tokens.
- anime optimises for booru-tagged models using tag-style tokens at higher density.
Pick the one that matches the base checkpoint you will train against.
Two shapes of the same image
The style does not change what the captioner sees — it changes how that gets written down. Toggle between them:
- What it does
- Natural-language tokens, phrased the way a realism checkpoint was captioned.
- Caption shape (example)
- sks woman, a woman in a charcoal wool coat standing on a rain-slick street at dusk, shallow depth of field, soft rim lighting
- Base checkpoints it suits
- SDXL / SD 1.5 realism checkpoints, Flux
Anime style also packs more in: a tag-trained checkpoint expects a denser token run, so the assembler allows more segments before it starts dropping them.
Picking one
Match the base checkpoint you will train against, not the subject matter. A
photograph captioned for Pony still wants anime style, because Pony was trained
on booru tags — the style describes the model's vocabulary, not the image's
content.
A quick test: if the checkpoint's own docs show example prompts full of
comma_separated_tags, pick anime. If they read like sentences, pick photo.
Where it shows up
Target Style in the Caption tool applies to every caption in the run. It
pairs with Target backend, which sets how many
of these tokens you can afford, and with
Target category, which picks which variant
becomes the final_caption.
Try it
- Open the Caption tool and expand the Target Style info panel — it reads from the same source as this page.
- Caption the same image both ways and compare the tag variant.