Prose enrichment
When enabled, novel noun/adjective phrases from prose output (Florence-2) are extracted and appended to the training variant as low-priority tag-style tokens.
This adds scene context without displacing core identity or wardrobe tags. Disable it for a pure WD14-only training caption.
What actually gets appended
Two captioners look at each image: WD14 produces tags, Florence-2 produces a sentence. Enrichment mines that sentence for phrases the tags missed and appends only those. Toggle it:
- Prose output (Florence-2)
- A woman in a wool coat stands on a rain-slick city street at night, lit by neon signage.
- Training caption
- sks womanwool coatcity streetnightrain-slickneon signage
- Appended, at low priority
- Only the novel phrases are added — woman, wool coat, city street, night are dropped because the tags already carry them. Being last, they are the first tokens a tight budget spends.
The two rules worth remembering:
- Novel only. A phrase the tags already carry is dropped, so enrichment never double-weights a concept it merely restated.
- Appended last. The new tokens land at the bottom of the priority order, so a tight token budget spends them first — they can never displace identity or wardrobe tags.
When to turn it off
Leave it on by default. Turn it off when:
- You want a pure WD14 tag caption — reproducing a tag-only dataset, or debugging which captioner is responsible for an artifact.
- You are training on a ~60-token backend and want the budget spent entirely on concept tags. (Though you can leave it on: the appended tokens are the first to be dropped anyway, so they cost you nothing you would have kept.)
Where it shows up
Prose Enrichment in the Caption tool applies to folder, Immich, and manifest runs. It is not accepted by the upload endpoint, so the toggle hides in Upload mode.
Do not confuse it with the tag ↔ prose balance control next to it: balance decides how much prose survives into the prose variant, while enrichment decides whether novel prose phrases are appended to the training variant.
Try it
- Open the Caption tool and expand the Prose Enrichment info panel — it reads from the same source as this page.
- Caption a folder with it on and off, then diff
final_captionagainstraw_prosein the raw outputs panel to see exactly which phrases were mined.