MiniMax H3高级配置·音频（Tlant）

MiniMax H3's headline trick is native, jointly-generated audio - the sound comes out with the picture, not bolted on afterwards. MiniMax H3高级配置·音频（Tlant） is the category that decides what that audio actually is: ambience, action sounds, background music, dialogue, language, and lip sync. It's one of the six config nodes feeding the Advanced Assembler, and the one with the most genuinely surprising options, because it's shaping a whole soundtrack, not just a vibe.

Same family architecture: typed 音频配置 output, plugs into the assembler's audio slot, assembler turns it into a stateless English instruction for a remote LLM. Unconnected, the assembler rolls seed-stable random audio - fine for experiments, but audio is worth controlling, because "random audio" is how you get a dramatic orchestral swell on a cooking video.

The inputs that matter:

• 音频模式 - the master switch: complete silence, ambient only, ambience + action sounds, music + ambience + action, or a full mix. 完全静音 tells the LLM to write N/A for both audio fields in the final H3 prompt.
• 音乐开关 - non-diegetic background music on, off, or auto. Distinct from 音频模式, so you can have ambience without music.
• 音乐风格 - cinematic score, epic orchestral, pop, electronic, TikTok beat, hip-hop, rock, ambient, jazz, piano/strings, acoustic, traditional. Theme inference is the safe default.
• 音乐速度 and 音乐能量 - tempo and energy curve (gradual build, peak-then-resolve, etc.). 音乐能量 pairs naturally with the action progression from the Action Options category.
• 节拍同步 - sync key actions, cuts, or both to the beat. This is the music-video magic knob.
• 动作音效 - physical sound density: fabric, footsteps, props. Sparse to rich.
• 环境声音 - the ambience layer: derive strictly from the visible scene, theme inference, or a rich layer.
• 对话模式 - no dialogue (the default, protecting face stability), one very short AI-written line, a short sung phrase, or auto. The tooltip is blunt: auto dialogue adds lip-sync and face-change risk.
• 对白语言 - only relevant if dialogue or singing is on: Chinese, English, Japanese, Korean, French, German, Spanish, or auto.
• 声音风格 - free text like "calm low female voice." Empty lets the LLM infer from the person and theme.
• 口型同步 - whether a speaking/singing subject must get matched lips.

Output: 音频配置, a custom typed port that only connects to the assembler's audio slot. Chinese labels, English final prompt.

Install

ComfyUI Manager → search ComfyUI-Tlant-Toolkit → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Tlant/ComfyUI-Tlant-Toolkit

Restart ComfyUI. No dependencies, no model downloads.

The gotchas

Dialogue is the trap of this category. Turn on 对话模式 and you're asking H3 to invent words, animate lips, and keep the face - the hardest combination in the pack. For realistic people, leave it on 无对话 unless dialogue is the entire point. Second, remember music and dialogue are separate toggles: you can have a full orchestral score with zero spoken words, and you should, because that's usually the better look. Silence is also a legitimate creative choice here - 完全静音 exists for a reason, and the node handles it cleanly by forcing both audio fields to N/A.