H3 Use Case Prompt Creator
Nine MiniMax H3 production recipes in one node — including the one for 'make a product ad'
- reference_image_1
- reference_image_2
- reference_image_3
- reference_image_4
- reference_image_5
- reference_image_6
- reference_video
- reference_audio
- h3_prompt
- production_brief
- use_case_analysis
The other two nodes in this pack solve the "I have an idea, write me an H3 prompt" problem. This one solves the "I have a brief" problem. MiniMax positioned H3 at commercial content creation - ads, brand promos, explainers - and its skills documentation reads like a production handbook: product ads want a variant, in-frame copy, a template and three independent anchors; co-op game intros want player names, a game title and a style lock; music videos want typography direction and beat-reactive shot structure. Knowing which details matter for which deliverable is exactly the kind of expertise a beginner doesn't have. This node encodes it.
Pick one of nine presets and the node shows you only the fields that use case actually cares about:
- Minimalist Product Ad, Brand Promo Video, 3D Animation Short, Co-op Game Intro, Handdrawn Live-Action Fusion, Music Video / Lyric Typography, Paper Collage Explainer, Papercraft Stop-Motion Explainer - or Custom H3 Use Case if none fit.
How it works
Each preset is a guide with its own required fields, production summary and reference rules. The rules are opinionated on purpose - "character refs are identity-only: face silhouette, hairstyle, proportions. Do not inherit photo realism, original lighting or source style," or the product ad's flat "do not use a 4-panel anchor sheet." That's the author's take on what the official skills demand, and it keeps a beginner from feeding the model a reference it will then blindly copy. The node collects only the filled fields for the active use case (stale values from a previous preset can't leak into the prompt), samples up to four frames from a connected reference video, and runs the whole thing through the same H3 guide engine as the rest of the pack - locally on Ollama or against a hosted provider.
The output is where this node stands out. You get three strings:
- h3_prompt - the actual guide-compliant three-field prompt for the sampler.
- production_brief - the planning document: truth sheet, story spine, beats, before any prompt exists. This is the "show your work" output that makes the workflow auditable.
- use_case_analysis - the reasoning behind it.
So it's not just a prompt writer; it's a tiny pre-production department. For a brand promo, for instance, it wants brand_name, official_url, campaign_focus, audience, cta, claims - and its guide explicitly refuses to invent logos, UI or metrics that aren't in your inputs.
The inputs
use_case (required), idea, duration (default 15s), aspect_ratio (16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or "match reference"), language, extra. Then the preset's optional fields light up - product_name/variant/copy/template for ads, story/character_count/world_style for 3D shorts, topic/learning_goal/visual_metaphor for explainers. You can also connect up to six reference images, a reference video and reference audio. There's a Built-in deterministic provider that skips the model entirely and returns a structurally valid (if boilerplate) prompt - handy when you just want to see the format before committing to a model call.
Installing
Same pack as the other two - ComfyUI Manager, search "MiniMax H3 Prompt Creator", restart, or:
cd ComfyUI/custom_nodes
git clone https://github.com/AIJigyasa/ComfyUI-H3-Prompt-Creator
No Python dependencies. The local path needs Ollama with a vision model:
ollama pull qwen3-vl:8b
Gotchas
Two things trip people up. First, the preset fields are dynamic - the widgets shown change with your use_case, which surprises people who expect every field all the time. Second, this node is the strictest of the three about required inputs: pick Co-op Game Intro with empty player1/player2/game_title/style and it will tell you what's missing instead of improvising. Treat that as a feature - H3's commercial formats punish hallucinated brand details. And as with the rest of the pack, keep temperature low (0.25 default) for schema adherence and num_ctx at 8192, or Ollama's default context sizing will drag your generation to a crawl.
Inputs (64)
| Name | Type | Default | Description |
|---|---|---|---|
| use_case | COMBO | Minimalist Product Ad | 9 options: Minimalist Product Ad, Brand Promo Video, 3D Animation Short, Co-op Game Intro, Handdrawn Live-Action Fusion, Music Video / Lyric Typography, +3 |
| idea | STRING | Core idea / goal | |
| duration | FLOAT | 15.01–120 | — |
| aspect_ratio | COMBO | 16:9 | 7 options: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, +1 |
| language | STRING | English | — |
| extra | STRING | Extra constraints / creative notes | |
| provider | COMBO | Ollama (Local) | Ollama (Local) needs no key. The hosted providers need api_key + api_model. |
| api_key | STRING | Blank reads OPENAI_API_KEY / ANTHROPIC_API_KEY / OPENROUTER_API_KEY / GEMINI_API_KEY. A key typed here is saved into the workflow JSON. | |
| api_model | STRING | Model for hosted providers; ignored by Ollama. Must be a vision model when reference images are connected. | |
| ollama_url | STRING | http://127.0.0.1:11434 | — |
| ollama_model | COMBO | qwen3-vl:8b | 1 options: qwen3-vl:8b |
| temperature | FLOAT | 0.250–1 | — |
| request_timeout | INT | 30030–3600 | — |
| num_ctx | COMBO | 8192 | Ollama context window. Leaving this to Ollama's default forces the KV cache off the GPU and makes generation extremely slow. |
| max_output_tokens | INT | 4096256–12000 | — |
| reference_image_1opt | IMAGE | — | |
| reference_image_2opt | IMAGE | — | |
| reference_image_3opt | IMAGE | — | |
| reference_image_4opt | IMAGE | — | |
| reference_image_5opt | IMAGE | — | |
| reference_image_6opt | IMAGE | — | |
| reference_videoopt | VIDEO | — | |
| reference_audioopt | AUDIO | — | |
| product_nameopt | STRING | — | |
| brand_nameopt | STRING | — | |
| official_urlopt | STRING | — | |
| campaign_focusopt | STRING | — | |
| audienceopt | STRING | — | |
| ctaopt | STRING | — | |
| claimsopt | STRING | — | |
| variantopt | STRING | — | |
| copyopt | STRING | — | |
| templateopt | COMBO | White-tech | 4 options: White-tech, Dark rim-light, Brand color field, Light lifestyle scene |
| storyopt | STRING | — | |
| character_countopt | INT | 21–20 | — |
| world_styleopt | STRING | — | |
| endingopt | STRING | — | |
| dialogue_styleopt | STRING | — | |
| player1opt | STRING | — | |
| player2opt | STRING | — | |
| game_titleopt | STRING | — | |
| styleopt | STRING | — | |
| ui_copyopt | STRING | — | |
| menu_themeopt | STRING | — | |
| contact_objectopt | STRING | — | |
| moodopt | STRING | — | |
| spaceopt | STRING | — | |
| drawn_entityopt | STRING | — | |
| language_constraintopt | STRING | — | |
| music_or_lyricsopt | STRING | — | |
| genreopt | STRING | — | |
| vocal_modeopt | STRING | — | |
| platformopt | STRING | — | |
| typography_directionopt | STRING | — | |
| character_directionopt | STRING | — | |
| reference_rolesopt | STRING | — | |
| topicopt | STRING | — | |
| learning_goalopt | STRING | — | |
| visual_metaphoropt | STRING | — | |
| paper_styleopt | STRING | — | |
| narration_languageopt | STRING | — | |
| paper_materialopt | STRING | — | |
| diorama_styleopt | STRING | — | |
| lesson_languageopt | STRING | — |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| h3_prompt | STRING | — |
| production_brief | STRING | — |
| use_case_analysis | STRING | — |