XB-llama - ✨ MiniMax-ref2va提示词预设

MiniMax H3 has a Ref2VA mode that's built for one thing: generate an audio-video from multiple reference clips - a reference image for the look, a reference video for the motion, a reference audio for the voices and sound. It's the mode you want for "make a new scene with my character, using this footage for the style." The catch: you have to tell the model what each reference is for, and how the styles should interact. XB_llamaMiniMaxRef2va builds that system prompt from a handful of dropdowns and text boxes - a sibling of XB_llamaMiniMaxPreset specialized for the reference-based mode.

Same architecture, different job: it outputs a system_prompt string for XB_llamaInstruct. No API, no keys, no weights - it's prompt engineering in a node.

The inputs

• 预设模式 - Ref2VA in English or Chinese.
• 参考图片介绍 / 参考视频介绍 / 参考音频介绍 - plain-text descriptions of each reference ("image one is a close-up of the male lead, video one is the camera-movement reference, audio one is the male voice"). This is where you tell the LLM what each reference contributes - and the node can't know, so this is genuinely on you.
• 视频时长 (4–15s), 画面比例, 切镜次数, 音乐风格 - the same parameter block as the fl2va preset, with the same timestamp/format enforcement.
• 视觉风格 - here it's different from the fl2va preset: just three strategies instead of 39 art styles:
  • 保持统一风格 / Consistent Style - everything rendered in one unified style from the references.
  • 多种风格混搭 / Mixed Styles - a live-action person can share a frame with a 2D character, each element kept in its own reference style (the prompt stresses stability: a real person must stay real throughout).
  • 多种风格转换 / Style Transformation - the whole frame smoothly transitions between styles over the duration (live-action gradually becoming watercolor, etc.).

Output

One system_prompt wire → XB_llamaInstruct.system_prompt. Then your own shot description goes in custom_prompt.

Installing it

Standard XB_ToolBox install - ComfyUI Manager (XB_ToolBox) or git clone, restart. No dependencies for the node; the llama stack (llama-cpp-python + a model in models/LLM) is needed to use its output. And remember the H3 license caveat from the fl2va article: the open weights are territory-restricted out of the US/EU/UK/Korea.

The judgment calls

The three style strategies are the whole point, and they're mutually exclusive by design - pick the one matching your actual goal. "Consistent Style" is the safe default and the right answer for character-driven scenes; "Mixed Styles" is the party trick (anime character interacting with a live-action person) and it's genuinely good at keeping each element stable; "Style Transformation" is the hardest to pull off and the most likely to melt into mush on a small model, so budget some retries. The reference descriptions are where people undershoot - a one-word note like "character" gives the LLM nothing to anchor on. Describe each reference's role concretely; the quality of the whole output tracks the quality of those three boxes.