XB-llama - ✨ MiniMax-fl2va提示词预设

MiniMax H3 is the 33B omni-modal video model that generates clips with audio - and its prompt format is brutal to hand-write. Every shot needs a [Shot N] At MM:SS.mmm timestamp, a style, a music directive, a non_diegetic_music field that must literally say "N/A" when you don't want a score... it's a format the model will punish you for getting sloppy with. XB_llamaMiniMaxPreset assembles that whole system prompt for you - pick the mode, the duration, the visual style, the music, the aspect ratio, and the cut count, and it emits a ready-to-use prompt the LLM can expand into shot-by-shot output.

It's a prompt builder, not the model itself: the node returns a system_prompt string that you wire into XB_llamaInstruct. No API keys, no network calls, no H3 weights on your disk - just an extremely opinionated prompt template.

The mode dropdown (预设模式)

Eight modes, covering the four input types H3 accepts, each in English and Chinese:

• T2VA - pure text-to-audio-video (no reference frames).
• I2VA - first-frame image to audio-video.
• FL2VA - first and last frame to audio-video (the pack's namesake - the display name literally says "fl2va提示词预设").
• L2VA - last-frame image to audio-video.

Each mode swaps in the matching base template, then injects your parameters.

The parameters it bakes in

• 视频时长 (4–15s) - H3's supported range. The prompt enforces exact duration and that the final shot ends before the second mark.
• 视觉风格 - 39 options, from cinematic to claymation to gongbi painting to ukiyo-e. Each carries a real style hint, and several use "GLOBAL MATERIAL OVERRIDE" language - the node genuinely understands that everything in frame must be clay, not just the character.
• 音乐风格 - 22 options (piano, orchestral, synthwave, guqin, Chinese opera...) with matching music descriptions, plus the "No Music" default that hard-forces non_diegetic_music to "N/A".
• 画面比例 - 8 ratios (16:9 → 5:4), each with composition guidance.
• 切镜次数 - 0–9 cuts, with a sanity check: if you ask for more cuts than the duration reasonably allows, the prompt tells the LLM to trim to at least 2 seconds per shot.

Output

One system_prompt string, built by splicing the parameter block into the mode's base template. Wire it into XB_llamaInstruct's system_prompt, feed your story in custom_prompt, and the local model generates the shot-by-shot script.

Installing it

Part of XB_ToolBox - ComfyUI Manager (XB_ToolBox) or:

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git

restart. The node itself needs nothing; to use the output you need the llama stack (llama-cpp-python + a model in models/LLM).

Things to know before you lean on it

The H3 weights come with a community license that excludes the US, EU, UK and Korea from the Applicable Territory - so "I'll run this locally" may not be legal where you are; the hosted Hailuo API remains an option. On the mechanics: the style hints are the real value here, but they're aggressive, and the "Material Override" styles will steamroll anything your own prompt says about the world being realistic. If you want a clay character in a real room, pick a non-override style instead. And the quality of the final script still depends on your local LLM following the format - an 8B model at high temperature will drift out of [Shot N] framing, which is exactly why the paired XB_llamaParameters defaults (temperature 0.6, repeat_penalty 1.12) exist.