Signs to Ideogram Caption

SIGN_DATA -> Ideogram 4 caption JSON. Wire `caption_json` into Ideogram4PromptBuilderKJ's `import_json` and set `import_mode` to 'always' — the builder's own elements_data is re-serialised from the browser at queue time, so any other route is silently overwritten.

By ping1979ping·Created 6 months ago·Updated 2 days ago· 1

sign_data
image

caption_json
box_prompts
mask
bboxes
width
height
report

◄high_level_description►

◄background►

◄lookphotographic►

◄text_inset0.07►

◄mask_pad_px4►

◄empty_too_smalltrue►

◄min_word_px0►

◄fallback_text►

◄auto_scenetrue►

◄lm_base_urlhttp://localhost:1234/v1►

◄lm_model_idqwen3-8b-vl-instruct-abliterated►

◄surface_hint►

CategoryFVMtools/signs

Inputs (14)

Name	Type	Default	Description
sign_data	SIGN_DATA		From Sign Text Proposer (or the selector, if you supply the words yourself).
image	IMAGE		The image the selector scanned. Only its size is used, to convert boxes into Ideogram's 0-1000 grid.
high_level_description	STRING		One line about the whole picture. Blank is allowed.
background	STRING		The scene around the lettering: light, weather, materials. This is what keeps the replacement in the same photograph as the rest of the frame.
look	COMBO	photographic	Ideogram drifts towards poster art unless told otherwise. Use 'flat graphic' only when the source genuinely is one.
text_inset	FLOAT	0.070–0.45	How far the TEXT box sits inside the region. The mask stays full size. A box that reaches the edge of a surface puts letters past that edge.
mask_pad_px	INT	40–64	Extra mask around each region. Err high: what the mask misses is restored unchanged and stands there as leftover scribble.
empty_too_small	BOOLEAN	true	Regions the selector judged too small for a readable word are described as blank instead of being written on. Off: every region gets a word, which produces letter-like shapes that count as invented text.
min_word_px	INT	00–200	Box heights below this get no word — the surface is described as blank instead. 0 turns the test off. The selector has its own height floor, but it measures the region BEFORE the line splitter cuts it into bands, so a 90px sign split into four passes the selector and arrives here as four 20px strips that the splitter only checks against its own 8px 'is this a line at all' bar. This is the last station before the caption and the only one that sees the geometry that will actually be rendered.
fallback_text	STRING		Used for a region the proposer left without a word. Blank means that region is emptied.
auto_sceneopt	BOOLEAN	true	Look at the picture and write the scene description itself when the two text fields above are empty. That is what lets the graph run from nothing but a chosen image.
lm_base_urlopt	STRING	http://localhost:1234/v1	LM Studio endpoint for the scene description.
lm_model_idopt	STRING	qwen3-8b-vl-instruct-abliterated	Name it. 'Whatever is loaded' is not a default, it is a silent failure: LM Studio lists models it has not loaded, the call returns nothing, and the run continues with an empty description. Must be a VISION model.
surface_hintopt	STRING		Per class, one per line, 'class: description'. Overrides the built-in wording for that class.

Outputs (7)

Name	Type	Description
caption_json	STRING	—
box_prompts	STRING	—
mask	MASK	—
bboxes	BOUNDING_BOX	—
width	INT	—
height	INT	—
report	STRING	—