DiffusionGemma Target Profile (Legacy — all models)

Every node in the DiffusionGemma Prompt Builder pack has a job, and this one is the quietest: it's a settings envelope. You tell it what you're generating for - an LTX video or an Ideogram 4 image, with sound design, duration, style, and negative-prompt policy - and it hands that contract to the CoT Generator so the model writes to the right target. No model inference, no pixels, no drama. It's the node you configure once and forget.

That's not to say it's dumb. The interesting part is how opinionated it is. Pick ltx and the model is instructed to return an empty Ideogram prompt; pick ideogram4 and the LTX prompt comes back empty, with scene segments collapsed to an empty list. Each profile gets its own output shape, and this node decides which shape.

How it works

Building the profile produces a DG_TARGET_PROFILE_CONFIG object that gets serialized into the model's prompt as "target controls." In the source that means a strict contract: target profile, audio mode, duration, style strings, aspect ratio, negative-prompt mode, and guidance all get embedded into the generation prompt, so DiffusionGemma knows the rules before it writes a word.

The negative-prompt policy is worth calling out because it's enforced twice - once in the prompt instructions, once after generation in the post-processor. auto tells the model to write a short negative only when it'll actually help; empty forces an empty string; custom returns your negative_prompt_guidance text verbatim and forbids the model from adding to it. So "trust me, this model wrote a weird negative" isn't a thing that happens here - the policy is a contract, not a suggestion.

The inputs that matter

• target_profile - ltx (default) or ideogram4. This is the big switch. Everything else on the node is profile-specific and ignored when it doesn't apply.
• audio_mode - auto_scene_audio (default) weaves one scene-specific sound cue into the LTX prompt when useful; explicit_sound_design tells the model to invent concrete diegetic sound from the visible action; visual_only bans audio cues entirely. LTX 2.3 generates synchronized audio, so this actually lands in your final video.
• target_duration_seconds - leave at 0 and the model uses the source video's duration (read from the Context Hub). Set it explicitly when you're generating from scratch or want a different pacing.
• ideogram_exact_text - for the ideogram4 profile only: literal text that must render verbatim. This is the field that makes Ideogram worth using over a generic image model.
• negative_prompt_mode - see above. Default auto is right for most people.

There's also ltx_style (a style hint prepended to the LTX prompt), ideogram_aspect_ratio (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3), ideogram_render_style, ideogram_json_output (whether Ideogram prompts come back as structured caption JSON), and audio_guidance.

Outputs are target_profile_config (the DG_TARGET_PROFILE_CONFIG you wire into the CoT Generator) and target_profile_json - a pretty-printed string if you want to see exactly what contract you just signed.

Install

Standard pack install: ComfyUI Manager (search DiffusionGemma Prompt Builder) or

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

then restart. This node needs the same dependencies as the rest of the pack (transformers>=5.12.1, comfy-kitchen, accelerate) and the DiffusionGemma NVFP4 model in ComfyUI/models/LLM/ before the pipeline does anything useful. Worth remembering while you're configuring: LTX 2.3 and Ideogram 4 are both community-licensed with real restrictions - LTX-2's license has a $10M revenue gate, and Ideogram 4's non-commercial agreement even reaches the outputs you generate. If commercial use matters to you, read those before you build a business on this.

Common issues

Honestly, the usual failure here is subtle and self-inflicted: forgetting you switched profiles. You tuned the Ideogram fields, left target_profile on ltx, and the model dutifully returns an empty Ideogram prompt because the contract told it to. If a target's outputs come back blank, check the profile switch first - and remember the splitter downstream blanks everything if ready_for_generation is false anyway.