MiniMax H3 Simple Prompt

H3 doesn't take a plain text box. It wants a structured prompt payload - text plus a mode plus a frame count plus a ratio - because one DiT needs to know whether it's doing text-to-video, image-to-video, or full reference generation before it samples. SimplePrompt is the honest name: it's the shortcut for people who just want to type a line and go, without touching the Storyboard editor or a refiner. It's the node you reach for on your first H3 render.

How it works

The node wraps your text and a handful of settings into a MINIMAX_H3_PROMPT dict: text, mode, frame_count, total_duration, ratio, and an optional negative_prompt. The frame_count is computed from total_duration at 24 fps, aligned to H3's frame-count convention. That dict is exactly the shape Conditioning and the Refiner nodes expect, so SimplePrompt is interchangeable with a Storyboard or a Refiner's output wherever a MINIMAX_H3_PROMPT is consumed.

The inputs that matter

• text - the actual prompt, multiline. Your shot description, in plain English. H3 is a vision-language-conditioned model; it reads this sentence by sentence, so describing camera, subject and motion beats keyword soup.
• mode - dropdown: T2VA (text to video+audio, the default), I2VA (image to video+audio), FL2VA (first/last-frame anchored), L2VA (last-frame only), full_reference (driven by reference media in a PackageData node). If you plug reference media or an FL constraint into Conditioning, the mode gets overridden by what's actually connected - so don't overthink it; set T2VA and let the wiring decide.
• total_duration - seconds, 1–15, default 5. This is your clip length; H3's max is 15 seconds.
• ratio - adaptive, 21:9, 16:9 (default), 4:3, 1:1, 3:4, 9:16. Conditioning's canvas logic will cap the actual pixel area at H3's 768×1344 limit, so tall/narrow ratios get normalized to whatever fits.
• negative_text (optional) - a negative prompt attached to this conditioning. Only bites if the KSampler's cfg is above 1.0.

The outputs

Two: prompt (the MINIMAX_H3_PROMPT payload, for Conditioning) and text (a plain STRING copy of the prompt, handy for display or a preview node).

Installing it

Pack-wide routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. It's a pure text node - no models needed.

Common issues

• Duration seems ignored. total_duration sets frame_count, and Conditioning builds the latent from it. If you changed the widget after wiring and nothing updated, drag the value; if you're feeding a Refiner in between, the Refiner carries its own duration and may override.
• Mode doesn't match what you intended. As designed - the connected FL constraint or package wins. Check Conditioning, not this node.
• Negative prompt does nothing. cfg is at 1.0. Bump it above 1.0 in the KSampler to activate negative guidance.

The one-liner: SimplePrompt + Conditioning + KSampler + Decode AV is the smallest H3 workflow that produces sound and video, and this node is the only part of it where you actually type.