H3 Auto Prompt Generator

MiniMax H3 is a genuinely good video model, and its prompt format is a genuinely tall order: six lowercase sections, reference labels, retention markers, shot timestamps in a specific syntax, and a mandatory task-type prefix - all of it judged against MiniMax's published writing guide. Writing that by hand for a character-replacement shot is doable. Writing it for every shot, every run, is a job you want a VLM to do.

H3 Auto Prompt Generator is that VLM. It reads a source video plus an identity reference image, picks keyframes on scene cuts and motion peaks, writes a production-ready REF2VA prompt in the official six-section format, then validates and repairs its own output - checking shot times, <Subject 1> tagging, wardrobe mentions, the summary prefix, retention markers, and a character budget - and retries once with the validator's error list. The "auto" in the name isn't marketing; the node is explicitly built so the model's mistakes get caught and fixed before the prompt reaches the sampler.

What you'll actually set

The required inputs are the personality of the prompt: reference_image (identity + wardrobe, <Picture 1>), subject_name (default "Aria Voss" - this becomes <Subject 1> throughout), subject_wardrobe (named 3+ times in the prompt, per the format), and scene_style (one line for detailed_description). Then vlm_provider picks the brain: anthropic, gemini (the only provider that can hear the audio track), openai, kimi, glm, qwen_api, or local options (qwen_local, minicpm_local, magevl_local, ollama). API keys read from env vars (or the optional api_key input).

Three toggles shape the output more than anything else:

• cut_times - paste or wire any Cut Detective output here. The cut list becomes ground truth: keyframes land on real shot starts, and the assembler forces the [Shot N] times onto the measured cuts instead of rescaling proportionally. A shot-count disagreement becomes a retry error, not a silent repair. This is the input that makes the pack's Cut Detective + H3 pairing coherent.
• music_video - writes the prompt as a music video: non_diegetic_music becomes the lead audio section and can no longer be N/A, cuts land on the beat, sung words go in lyrics (written as <d>[Language] ...</d> in the shot where they're heard) and the track in music_description. Leave lyrics blank and H3 invents nonsense syllables - the tooltip says it plainly.
• first_frame_alignment - the I2V hook. On, <Picture 1> becomes the literal opening frame; off (the default), it's an identity-only reference whose background must not leak into the video. This is the one people get backwards: on for image-to-video, off for character replacement.

The format details that make it trustworthy

The README is refreshingly specific about what "official" means here: every rule the assembler enforces traces to MiniMax's own VIDEO_PROMPT_WRITING_GUIDE_ref_en.md and ..._base_en.md, and tests/test_h3_format.py asserts each rule against the quoted text. Two things stand out - the trailing block of "No ..." exclusion sentences is not in the official format (it's off by default behind append_exclusions, because H3 has no negative-prompt field), and non_diegetic_music: N/A genuinely means "there is no non-diegetic music."

Outputs: h3_prompt (and h3_prompt_b in both_ab profile mode for A/B testing), duration_seconds, fps, and frame_analysis_json.

Gotchas

• This is a hosted-API node by default: it needs a provider key, and gemini is the only backend that can describe what's actually heard in the audio. The model field defaults to auto per provider.
• prompt_profile matters: official is strict HF-guide format; upgraded adds battle-tested practices (positive assertions, ~3000-char budget, camera speed words); both_ab runs two VLM calls so you can compare.
• The validator retries once with its error list - so a run can take two API calls. That's a feature (the format gets repaired), just budget for it.

Install

One of ~69 nodes in TrentNodes. The pack install gets you the node; the VLM providers need their Python clients (anthropic, openai, google-genai are in the pack requirements) and keys:

# ComfyUI Manager: search "Trent Nodes"

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes && pip install -r requirements.txt

No local model downloads unless you pick a local provider. If you've been hand-writing H3 REF2VA prompts, this is the node that makes the format legible instead of a chore.