Target backend

The diffusion model architecture that will consume these captions. Each backend has a different CLIP/T5 token budget — exceeding it means wasted tokens the model never sees.

SDXL / SD 1.5 / Playground: ~60 tokens.
Flux / SD3 / Kolors / PixArt: ~200 tokens.

Captions are assembled to fit the selected backend's budget, dropping the lowest-priority fragments first.

Spend the budget

Pick a backend and watch the same caption fit — or not. The budgets are the ones the Caption tool sends, so this is the trade you are actually making:

Try itbudgets from the Caption tool

sdxl gives you 60 tokens. This example caption wants 160 — the 3 lowest-priority fragments are dropped.

trigger word + identity tags18t
wardrobe / garment tags22t
pose and framing20t
dropped: setting and background34t
dropped: lighting and mood26t
dropped: appended prose tokens40t

Fragment sizes are illustrative; the budgets are real. Ordering is by training priority — identity survives, prose is spent last.

The lesson is priority, not length. A caption is assembled highest-priority-first, so a tight budget does not truncate mid-sentence — it drops whole low-priority fragments. Identity survives; appended prose is spent first.

Choosing one

Match the backend to the checkpoint you will train against, not to the caption you wish you could write:

~60 tokens (SDXL, SD 1.5, Playground) — budget is scarce. Lean on tags over prose and let the low-priority fragments go.
~200 tokens (Flux, SD3, Kolors, PixArt) — room for scene context, so prose enrichment earns its place.

Exceeding the budget is not an error; it is waste. The tokens past the limit are assembled, then never seen by the model.

Where it shows up

Target Backend in the Caption tool sets the budget for every caption in the run, single or batch. It pairs with Target style, which decides the shape of the tokens the budget is spent on.

Try it

Open the Caption tool and expand the Target Backend info panel — it reads from the same source as this page.
Caption one image at ~60 tokens, then again at ~200, and diff the final_caption.