DiffusionGemma MiniMax-H3 Target Profile

MiniMax H3 is the omni model that renders a video and its audio in one pass - but its prompt format is brutal to hand-write. It wants a strict [Shot N] timeline, soundscape and score separated into their own sections, and for Ref2VA a six-section reference contract with <Subject N> semantics. DiffusionGemma exists to translate your rough brief into exactly that structure. DiffusionGemma MiniMax-H3 Target Profile is the node where you tell the translator which H3 dialect you're speaking and how much of it you want: T2VA or Ref2VA, duration, shot count, dialogue rules, audio policy.

It's one of the three "choose exactly one" Target Profiles (LTX-2.5, H3, Ideogram 4). In the H3 wiring, its target_profile_config output feeds both the CoT Generator and the JSON Splitter - the splitter needs it too, because it decides how to unpack Director's JSON into the H3-native prompt.

The inputs that decide your prompt

• generation_mode - t2va writes H3's canonical three-field text-to-video prompt (integrated_multimodal_description, overall_soundscape, non_diegetic_music). ref2va writes MiniMax's full six-section reference contract (subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music) for MiniMaxH3ReferenceToVideo. Ref2VA needs the H3 Reference Context upstream with a real image/video manifest.
• target_duration_seconds - 0 auto-resolves from source media, then a duration in the brief, then a 5-second default. MiniMax documents 4–15s as H3's happy range, and the README treats that as quality guidance, not a clamp - you can request 20s and let your VRAM, sampler, and render time decide whether it survives.
• shot_count - auto honors an exact count written in your brief, otherwise Director picks the fewest useful shots. Or pick 1–12, or custom with a value up to 99 in custom_shot_count. A shot is a camera/cut segment, not a generation pass - several native [Shot N] blocks can live inside one valid 4–15s H3 clip. Don't confuse this with a Project Master generation-lane count; that's a separate thing.
• dialogue_mode - auto preserves speech you explicitly asked for but never invents it; required authorizes Director to write dialogue and enforces the exact block count; off bans speech, narration, and singing while keeping ambience and music. dialogue_line_count (1–12, default 2) and dialogue_guidance only matter in required mode.
• audio_mode / audio_guidance - scene audio policy (auto_scene_audio, explicit_sound_design, visual_only) and direction for ambience, effects, music, or quoted speech. Note required dialogue is incompatible with visual_only and gets rejected before generation.
• negative_prompt_mode / negative_prompt_guidance - H3 has no separate negative-conditioning channel through its local text encoder. auto/custom fold relevant exclusions into the integrated description as natural language; empty ignores the guidance while preserving exclusions written in the brief.

How Director turns those into H3 prose

Every H3 shot carries one explicit camera-behavior sentence. Moving-shot prose names the motion type, speed/amplitude, direction, subject relationship, and ending frame; a locked-off sentence is only the deterministic last resort for a genuinely unspecified shot. Dialogue gets real markup - the speaker ID sits outside the block, language and words inside:

The protagonist (S1) says: <d>[English] We made it this far.</d>

The hard gate is structural: it verifies dialogue presence, exact block count, speaker cues, and native markup. It does not claim to prove your free-form dialogue guidance was semantically obeyed - so if you need exact lines, put them in dialogue_guidance as quoted wording.

Installing it

Part of exportAnything/ComfyUI-DiffusionGemmaPromptBuilder:

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

or ComfyUI Manager → DiffusionGemma Prompt Builder → restart. The config node itself is lightweight; what's not is the DiffusionGemma model folder at ComfyUI/models/LLM/diffusiongemma-26B-A4B-it-NVFP4 and, if you actually render, the H3 weights (pruned INT8/ConvRot Ref2VA diffusion model, Qwen3-VL text encoder, video + audio VAEs, Turbo LoRA) listed in the README.

Common issues

• Wrong shot count keeps blocking. An explicit shot_count requires exactly that many consecutive [Shot N] sections; Director repairs once, then the gate blocks. Counts above ~12 generally need more duration and output-token budget.
• Dialogue that doesn't fit. Twelve shots in 15 seconds is 1.25s per shot - no line survives that. Reduce shot count, dialogue count, or raise duration instead of forcing speech across rapid cuts.
• Ref2VA validation spam. Strict Grounding Guard demands [dg:...] annotations on every connected Picture/Video role in the manifest; without them, role validation fails deterministically. Annotate the manifest or run audit mode.
• Don't wire the Project Master's lane recommendation into shot_count. Native storyboard count and generation lanes are different axes; mixing them up is the most common wiring mistake on this node.