H3 Video Prompt Creator — idea + optional first/last frame. Automatically selects T2VA/I2VA/FL2VA/L2VA and writes the complete H3 three-field prompt.
H3 Full-Reference Video Prompt Creator — idea + optional IMAGE/VIDEO/AUDIO references. Builds a reference dossier, resolves <Subject N>, <Picture N>, <Video N>, <Audio N>, and writes the six-section full-reference prompt.
H3 Use Case Prompt Creator — specialized H3 workflows based on the official MiniMax H3 skills.

Source guides

The prompt rules are encoded from two official documents:

Video Prompt Writing Guide (T2VA / I2VA / FL2VA / L2VA)
Full-Reference Mode Rewrite Output Format Guide

They live as VIDEO_SYSTEM and FULL_REF_SYSTEM in h3_prompt_creator.py. Rules a language model reliably ignores are additionally enforced in code after generation: shot timing, reference-label numbering, subject de-duplication, and refusal to cite an asset that was never connected. If you revise the guides, update both the system prompt and the matching enforcement.

Providers

All three nodes can run against a local model or a hosted API. Pick with the provider widget:

| Provider | Needs | Default model | Env var | |---|---|---|---| | Ollama (Local) | nothing | qwen3-vl:8b (from ollama_model) | — | | OpenAI | key | gpt-4o | OPENAI_API_KEY | | Anthropic | key | claude-opus-5 | ANTHROPIC_API_KEY | | OpenRouter | key | anthropic/claude-sonnet-5 | OPENROUTER_API_KEY | | Google Gemini | key | gemini-2.0-flash | GEMINI_API_KEY / GOOGLE_API_KEY | | Built-in deterministic | nothing | — | — |

Prefer the environment variable over the api_key widget. A key typed into the widget is saved into the workflow JSON and travels with the file if you share or export the workflow. Leave api_key blank and the node reads the variable above.

api_model overrides the default. It must name a vision-capable model whenever reference images are connected. ollama_url / ollama_model / num_ctx / keep_alive apply only to the local provider.

Everything is plain HTTP through urllib — no vendor SDKs, so the node adds no pip dependencies. Provider-specific handling worth knowing:

Anthropic: temperature is omitted for models that reject it (Claude Opus 5, Sonnet 5, Opus 4.8/4.7, Fable 5 return a 400 if it is sent). Thinking is on by default on current Claude models and shares the max_output_tokens budget with the reply, so leave that generous. A stop_reason: refusal is reported rather than silently returning empty.
OpenAI: if the model rejects max_tokens, the request is retried with max_completion_tokens; if it rejects a non-default temperature, that is dropped and retried.
Gemini: uses responseMimeType: application/json for JSON replies.
A missing key fails immediately with a clear message, before any slow call.

Local AI pipeline

Qwen3-VL 8B via Ollama: image analysis, sampled-video visual analysis, reference reasoning, and final H3 structuring.
faster-whisper: local speech/lyrics transcription with timestamps.
librosa: local audio/music feature analysis.
FFmpeg: optional extraction of embedded audio from reference VIDEO sources.

The H3 nodes intentionally separate reusable subject identity from incidental source-image background/pose/camera/lighting unless the user explicitly asks to preserve those details. This reduces background/pose leakage when a reference is being used only for character identity.

Install

ComfyUI Manager — search for "MiniMax H3 Prompt Creator" and install, then restart ComfyUI.

Manually — clone into ComfyUI/custom_nodes/ and restart:

git clone https://github.com/AIJigyasa/ComfyUI-H3-Prompt-Creator

The node installs with no Python dependencies — the prompt engine talks to every backend over plain HTTP.

Audio analysis is opt-in, because faster-whisper and librosa are large. Without them the nodes simply report that transcription is unavailable; everything else works. To enable it, install into the ComfyUI environment:

pip install -r requirements-audio.txt

(The file is deliberately not called requirements.txt — ComfyUI Manager installs that automatically, which would make every install a heavy one.)

FFmpeg should be available on PATH for embedded audio extraction from reference videos.

Ollama

Install a vision-capable Qwen model, e.g.:

ollama pull qwen3-vl:8b

Default Ollama server: http://127.0.0.1:11434

The nodes fall back to deterministic H3 formatting if Ollama is unavailable, but reference/media understanding is naturally better with Qwen3-VL available.

Audio

Reference audio is never passed directly to Qwen3-VL. Instead:

faster-whisper creates transcript + timestamps for speech/lyrics.
librosa extracts duration, tempo estimate, spectral centroid/rolloff, zero-crossing and RMS dynamics.
FFmpeg can extract audio embedded in a reference VIDEO.

The resulting evidence is supplied to Qwen3-VL as structured text for H3 reasoning.

Troubleshooting Qwen3-VL on Ollama

This package uses Ollama's local HTTP API at http://127.0.0.1:11434. Qwen3-VL must be installed in Ollama and the installed name must match the node's ollama_model value (for example qwen3-vl:8b). The model dropdown is populated from your running Ollama server at ComfyUI start; use the Refresh Ollama Models button after pulling a new model.

Recommended checks in a terminal:

ollama list

If the node runs forever and produces nothing

Three separate causes were behind this; all are fixed in the node, but they are worth understanding because they affect any Ollama integration.

1. The context window (the big one). If a request does not set num_ctx, Ollama sizes the context from the model's maximum — 262144 tokens for Qwen3-VL. That needs roughly 25 GB of KV cache, so on a 12 GB card most layers spill to system RAM. Check with:

ollama ps

A healthy row reads 100% GPU. If it reads something like 62%/38% CPU/GPU, generation drops from ~50 tok/s to ~5 tok/s and a single prompt can take 15 minutes. The num_ctx widget (default 8192) pins this. It auto-raises in fixed steps when many reference images are connected, so the model stays loaded instead of reloading on every run.

2. The answer arrives in the wrong field. On current Ollama builds, Qwen3-VL returns its reply in message.thinking and leaves message.content empty — even with "think": false. Code that reads only message.content gets an empty string every time and falls back to boilerplate. The node now reads whichever field is populated.

3. Truncated JSON. When generation hits num_predict, the JSON object is cut off mid-string and normal parsing discards the whole response. The node repairs unterminated JSON so a truncated generation is still usable, and reports clearly when it cannot.

If you get dozens of near-identical subjects and no actual prompt

A reference that shows the same person more than once — a multi-pose contact sheet, a turnaround, a collage — makes the vision model treat each pose as a separate person. It then restates the same description until the token budget is gone, so subject_definitions fills with 27 copies of one woman and detailed_description is never written.

Each reference asset is analyzed in its own request. Sending every image in one batch makes the model cross-attribute — describing image 2's person under <Picture 1>. One asset per request removes the ambiguity, at the cost of one model call per connected reference. Turn off enable_reference_analysis if you would rather have the speed.

A character sheet counts as one subject. The analyzer is told that the same person in different poses, angles, crops or outfits is a single subject, described as a whole person with the outfit variants noted inside that one definition — not one subject per panel.

Further defences, in order:

The analyzer is told that one person is one subject regardless of pose, and that "left/middle/right" are not different people.
The schema caps subjects (3 per asset, 6 global) and the analyzer gets its own smaller token budget, so a runaway dossier cannot eat the whole run.
Near-identical subjects are merged in code after generation, ignoring positional wording. Genuinely different people are preserved. The merge count appears in generation_notes.
If summary, retention_analysis or detailed_description still come back empty, the node retries once, naming the empty fields and capping the subject list.
Reference labels are renumbered across the whole prompt so <Subject A> becomes <Subject 1> and every section agrees.
Shot timestamps that fall outside the target duration are respaced inside it, and Shot 1's timestamp is stripped.

Speed reference

On an RTX 3060 12 GB with qwen3-vl:8b at num_ctx 8192:

| Operation | Time | |---|---| | Text-only prompt (T2VA) | ~5 s | | One reference image (I2VA) | ~15 s | | Three reference images, full-reference node | ~50 s |

If runs are much slower than this, check ollama ps for CPU offload first.

Other controls

enable_reference_analysis — the nodes make two model calls: a reference-dossier pass, then the prompt itself. Turning this off halves generation time when references are simple.
max_output_tokens — raise if generation stops mid-sentence; the node reports when it hits the limit.
Reference images are downscaled to 768px on the long edge before being sent, since Qwen3-VL's vision cost scales with resolution.