Ideogram4-KoboldCPP

Ideogram 4 doesn't take a plain prompt - it takes a structured JSON object with a high-level description, a background description, and a list of bounding-boxed elements. Editing that by hand is miserable, and rewriting the whole thing in one LLM call bleeds details between regions. Ideogram4-KoboldCPP fixes both: it splits the JSON into fragments, rewrites each one in a separate local KoboldCpp request, and reassembles the structure with its boxes and types untouched.

This is a niche node for a specific model. If you don't run Ideogram 4 weights (the 9.3B DiT whose whole claim to fame was best-in-class text rendering and JSON-only prompting - see the KB's ideogram-4 panel), you can stop reading here. If you do, this is the node that turns "make the background moodier" into a targeted, structure-preserving edit instead of a hand-edited JSON blob.

How it works

The node walks your JSON and rewrites these fields (each gated by a process_* switch, all on by default):

• high_level_description
• compositional_deconstruction.background
• compositional_deconstruction.elements[*].desc - except elements whose type is "text", which are preserved literally because you don't want an LLM "fixing" your in-image text.

Everything else - bounding boxes, element types, literal text elements, unknown keys, unselected fields - is passed through byte-for-byte. Crucially, the model only ever sees one prompt fragment at a time. It never sees the JSON structure, the field name, other regions, or the bounding boxes, which is how detail stays put instead of leaking across the image.

The switches that matter:

• process_high_level_description, process_background, process_elements - turn off any region you don't want touched.
• seed_mode - Offset per field gives each rewritten fragment its own seeded dice roll; Same seed makes every field use the same seed.
• on_error - Stop aborts the workflow if a fragment rewrite fails; Keep Original keeps the source fragment and continues.
• pretty_json - formatted vs. compact output. Keep it on for readability.

On top of that you get the full amplification control set: model_profile, style_preset, framing_preset, embellishment_level, thinking_mode, additional_instructions, and the usual KoboldCpp request knobs (URL, token allowance, temperature, penalties). Each fragment is rewritten using those same settings, so one style decision propagates consistently across the whole JSON.

The single output is json_output (STRING) - feed it back into whatever reads Ideogram 4's structured prompt.

Installing it

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/tiko13/ComfyUI_PromptStudio

Restart ComfyUI, or install "ComfyUI_PromptStudio" via ComfyUI Manager. No pip dependencies - you need a local KoboldCpp server (default http://localhost:5001; Ollama and llama.cpp work too), and for KoboldCpp specifically, enable Use Jinja and restart so the chat-template rewrite works.

Where people get burned

• Malformed input JSON. The node expects the Ideogram 4 structured-prompt schema. If you paste something else in, garbage out - it won't invent the schema for you.
• Expecting region coordination. Each fragment is rewritten blind. The node preserves structure on purpose; if you need the background to mention what's in the foreground, put that in additional_instructions.
• Text elements changing. They won't - by design. If you want your in-image text reworded, this node is the wrong tool.
• VRAM contention. Same as every KoboldCpp node here: on a single GPU the LLM and diffusion model share VRAM, and the shared-GPU handoff needs KoboldCpp Admin Mode + an Admin Directory. Multiple fragment rewrites mean multiple LLM calls per run, so give the handoff time to work.