Prompt builder
Compose across every axis of the catalog and get a prompt in the language your model family actually speaks. The three families do not share one, which is why this is not a single output box.
Nothing is generated on this page and nothing is sent anywhere. The builder composes text from the published vocabulary in your browser, so there is no account, no queue and no cost. Take the output to whatever you already run.
Where the catalog knows a mapping is imprecise, the builder says so rather than quietly emitting a tag that means something broader. Read the warnings under the output before you trust a prompt.
2 participants, receiving partner on top, face to face, on a flat surface.
Comma-separated underscored tags. Only tags the model saw in training do anything.
Positive
1boy, 1girl, hetero, amazon_position, girl_on_top, squatting, legs_up, on_back, straddling, on_bed
Negative
bad_anatomy, bad_hands, censored, monochrome, 2girls, 3girls, 2boys, 3boys
Before you run this
- Expected figure count is 2. Global pose conditioning is not known to hold figure count on multi-participant positions: count figures and limbs at full size before judging quality.
Amazon expects 2 figures. Folksonomy tag amazon_position (exact).
Why three outputs
Booru-trained checkpoints answer to underscored tag strings and ignore prose. Photoreal checkpoints want sentences and degrade on tag soup. Video models want a clinical description of motion and nothing else. One canonical prompt would serve none of them well.
Why the figure-count warning
Global pose conditioning does not make a diffusion model count bodies. On multi-participant positions it produces convincing images with the wrong number of limbs. Count figures at full size before judging quality: thumbnails hide it.
Where the tags come from
Every tag was checked against the live tag API for existence and deprecation, and the mapping to each position was verified against the tag's own documentation. The crosswalk records what each imprecise mapping discards.
Vocabulary generated 2026-08-09. Open data, CC BY 4.0. Full dataset on Hugging Face →