DiffusionGemma Ideogram 4 Target Profile

Ideogram 4 is the still-image fork of the DiffusionGemma Director family, and it exists because image prompting and video prompting are different jobs with different output contracts. The LTX and MiniMax-H3 profiles agonize over duration, shot timelines, camera moves, and audio policy. Ideogram doesn't care about any of that - it cares about aspect ratio, render style, and, above all, the exact text you want burned into the image. DiffusionGemma Ideogram 4 Target Profile deliberately has no duration and no audio widgets, which is the entire point: it tells DiffusionGemma to write an Ideogram-ready prompt for a single still.

Ideogram's claim to fame, and its weak point, is typography - it renders legible text in images better than almost anything open, and DiffusionGemma is good at making up believable-looking gibberish signage. exact_text is where you close that gap: put the wording you actually want rendered, verbatim. Leave it blank when the image should have no text, or the model may invent signage you then have to fix.

The inputs that matter

• aspect_ratio - 1:1 default, with 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 available. Note the splitter no longer inherits Ideogram's default into other models: LTX defaults frame-free to 16:9 independently.
• render_style - optional visual medium, finish, or art direction ("flat vector illustration", "painted poster look").
• exact_text - the visible wording Ideogram should render. The one field most people underuse.
• json_output - a representation switch, not an extra socket: on returns the ideogram_prompt as caption-schema JSON (for ComfyUI's native Ideogram loader), off returns ready-to-use prose. Toggle it for the path you're on.
• negative_prompt_mode - auto, empty, or custom (with negative_prompt_guidance), matching the other profiles.

Outputs are target_profile_config and target_profile_json, wired into the CoT Generator exactly like the other profiles. Choose one Target Profile between the Context Hub and the Director - never stack Ideogram and LTX profiles in one prompt-authoring chain.

How it fits the Director philosophy

DiffusionGemma reviews your source image and/or video and writes a structured Ideogram4 prompt through its CoT Generator, with the JSON Splitter producing the ideogram_prompt output plus metadata and resolution. Same five-node core, different target contract. The author's community posts show the Ideogram side working "just as well" as the LTX side for reviewing an image and producing a strong structured prompt - it's the natural still-image output of a graph that already does video.

One honesty note worth carrying in from Ideogram's own history (it's in the KB): Ideogram 4 ships a non-commercial license and quantized checkpoints only, and has a trained-in refusal-image behavior. So this node is for your own creative stills, not for a commercial pipeline, and the exact-text guarantee is about the prompt contract - Ideogram's own rendering quality and licensing constraints still apply downstream.

Installing it

It's one of 55 nodes in exportAnything/ComfyUI-DiffusionGemmaPromptBuilder:

cd ComfyUI/custom_nodes
git clone https://github.com/exportAnything/ComfyUI-DiffusionGemmaPromptBuilder

or search DiffusionGemma Prompt Builder in ComfyUI Manager and restart. Requirements are the pack-wide Python set plus the DiffusionGemma model folder at ComfyUI/models/LLM/diffusiongemma-26B-A4B-it-NVFP4. This node is pure configuration; it costs no extra VRAM of its own.

Common issues

• Text you didn't ask for shows up in renders. You left exact_text blank and DiffusionGemma filled the void - poster text, signage, watermarks are all things a multimodal prompt-builder loves to invent. Blank it intentionally, or feed the exact copy you want and treat that as authoritative.
• Wrong aspect on output. The splitter's resolution controls are downstream sampler controls; if your Ideogram image comes out the wrong ratio, check the splitter's resolution_aspect_ratio selection, not the target profile alone.
• Confusing the profiles. If your graph has an Ideogram profile and an LTX profile in the same CoT chain, behavior is undefined-ish and the gates get confused. One target profile per Director path.