Signs to Ideogram Caption

Ideogram 4 shipped its weights in June 2026 - 9.3B, non-commercial licence, quantized checkpoints only, and a safety filter so eager it emits a dedicated refusal image rather than your prompt. It was the best text-in-image model open weights had ever seen, and it talks in JSON, not prompts. That's the whole reason this node exists: it converts the pack's SIGN_DATA into the structured caption the Ideogram node actually consumes, so a chain that starts with "find the lettering and decide what it should say" can end in an Ideogram render. (It's also why the community's LoRA production moved to Krea 2 three weeks later - better renderer, worse licence terms, same old story.)

The one thing you must get right

Wire caption_json into Ideogram4PromptBuilderKJ's import_json input and set import_mode to always. That isn't a preference, it's the only route that works. The builder's elements_data widget is re-serialised from a live array in the browser every time the graph is queued, so anything written into it from outside gets overwritten the instant you press run - silently. If your caption seems to be ignored, this is why, and it's the first thing to check.

Inputs

• sign_data - from Sign Text Proposer, or from the selector if you're supplying the words yourself.
• image - only its size is used, to convert boxes into Ideogram's 0–1000 grid. Feed the same image the selector scanned.
• high_level_description - one line about the whole picture. Blank is allowed, but blank is worse.
• background - the scene around the lettering: light, weather, materials. This is what keeps the replacement inside the same photograph as the rest of the frame.
• look (photographic / flat graphic) - Ideogram's untended pull is poster art: ask for a sign and it answers with flat vector lettering on a clean plate. The caption therefore states the camera, not the design. Only switch to flat graphic when the source genuinely is one.
• text_inset (0.07) - how far the text box sits inside the region. The mask stays full size; the box shrinks. A box that reaches the edge of a panel puts letters past that edge.
• mask_pad_px (4) - extra mask around each region. Err high: a mask that misses a stroke leaves the old scribble standing bit-identical, which is the single most visible failure in this pipeline.
• empty_too_small (on) - regions the selector judged too small for a readable word are described as blank rather than written on. Turn it off and you get letter-like shapes where words should be, and Ideogram counts those as invented text.
• min_word_px (0) - the last safety net before the caption. Box heights below this get no word; the surface is described as blank instead. Leave it on if you're splitting multi-line signs, because the selector measures regions before the split.
• fallback_text - used for a region the proposer left wordless. Blank means that region gets emptied.
• Optional auto_scene (on) with lm_base_url / lm_model_id - if the two text fields above are empty, the node writes the scene description itself with the local vision model, which is what lets the graph run from nothing but a chosen image. surface_hint overrides the built-in wording per text class, one class: description line each.

Outputs

caption_json into the builder's import_json; box_prompts for reading what was decided; mask for whatever inpaints or re-renders the regions; bboxes, width, height for debug and for grid math; report for the log.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

Restart ComfyUI, or install from ComfyUI Manager with comfyui-fvmtools. No extra Python packages for this node - it's numpy, torch and json. Ideogram 4 itself is a separate download, you'll want Kijai's prompt-builder node alongside it, and remember that Ideogram's weights are non-commercial.

Where people get burned

import_mode left on its default is the number one cause of "the caption did nothing". Check that first, before you touch a single parameter.

Leftover scribble around the new text means the mask is too tight - raise mask_pad_px. Text spilling off the edge of a sign means the opposite problem, the text box is too big - raise text_inset.

And resist the urge to describe the design here. This node's caption fields are about the photograph; the lettering itself comes from the proposer. Fill the design in by hand and you'll fight the model, because Ideogram's guidance - confirmed by the author's own testing - is that describing the scene as photographic is the counterweight to its pull toward flat poster lettering.