JZL - ✨ MiniMax-ref2va提示词预设

Reference-to-video is where MiniMax H3 stops being a toy. You hand it a handful of character shots, a background, maybe a clip for motion or a voice to match, and the model builds the scene around your elements instead of inventing its own. JZL - ✨ MiniMax-ref2va提示词预设 is the prompt builder for exactly that path: it takes your description of the references, plus the same style controls as the standard preset, and assembles the ref2va system prompt for you.

It's the sibling of the fl2va preset (JZL_MiniMaxPreset), but the whole shape changes: instead of eight generation families you get two - 多参考生成音视频 [EN] and [ZH], multi-reference text-to-video-plus-audio in either language. Everything else is tuned for the "these references are my cast" workflow.

The inputs that matter

• 参考图片介绍 / 参考视频介绍 / 参考音频介绍 - plain-language notes about each reference: "图一男主角特写，图二女主角全身" (image one is a close-up of the male lead), "视频一运镜参考" (video one is a camera-move reference), "音频一男主音色" (audio one is the male lead's voice). The node injects these verbatim into the prompt so H3 knows what each slot is for. This is the input that decides whether your references get used as intended or as a confused mush.
• 视觉风格 - only three options, and they're the interesting part:
  • 保持统一风格 / Consistent Style - everything must look like one unified universe; nobody gets to look like they wandered in from another artwork.
  • 多种风格混搭 / Mixed Styles - a live-action actor next to a 2D-animated character is fine, but each element must stay visually stable: the anime character stays anime, the clay figure stays clay, for the whole clip.
  • 多种风格转换 / Style Transformation - the entire frame morphs from one style to another over the duration (live-action gradually becomes watercolor), while shapes and layout stay put.
• 视频时长 - 4–15s, and the node enforces it in the prompt: timestamps must land inside the window, last shot ends before the limit.
• 音乐风格 / 画面比例 / 切镜次数 - the same menus as the fl2va preset, including the sanity warning if you ask for more cuts than a short clip can carry.

Output and wiring

One output, system_prompt (STRING), which feeds the H3 prompt path. And if this pack's README headline is any hint, the intended pairing is its own reference encoder - the 🎬 MiniMax H3 参考编码 node (JZL_MiniMaxH3ReferenceToVideo) is a from-scratch replica of the official ref2va conditioning node with dynamic auto-growing reference slots. Reference images, videos, and audio go in there as <Picture i>, <Video k>, <Audio j> placeholders; the prompt text uses the same tags, so keep the two sides in sync - describe exactly the references you actually wired in.

Install and the usual caveats

Same pack install as everything else here: ComfyUI Manager → search JZL or ComfyUI-JZL-MiniMax-H3, or clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3 into custom_nodes/ and restart. Recent ComfyUI required (V3 io.Schema / official H3 nodes). No LLM runtime needed - this is pure string assembly.

Labels are Chinese-first with English in brackets, so 参考图片介绍 will trip you up until you've typed it twice. And the recurring H3 license reminder: the open weights are geofenced out of the US, EU, UK and South Korea, so check your region before building the ref2va workflow on local weights - the hosted API is the alternative.