Nodes/ComfyUI-Tlant-Toolkit/MiniMax H3高级配置·人物一致性(Tlant)
ComfyUI Node

MiniMax H3高级配置·人物一致性(Tlant)

Keeping the same face for ten seconds of video

By Tlant·Created 3 months ago·Updated 28 days ago· 0
MiniMax H3高级配置·人物一致性(Tlant)
    • 人物一致性配置
    身份锁定严格
    面部动作微表情
    头部动作轻微
    面部朝向保持原图
    视线行为保持原图
    表情变化保持原表情
    眨眼自然一次
    嘴部动作保持原状
    面部遮挡禁止

    If there's one thing that separates a good image-to-video generation from a horror show, it's the face staying the same face. MiniMax H3高级配置·人物一致性(Tlant) is the category node dedicated to exactly that - identity, facial motion, direction, gaze, expression, blinks, mouth, and occlusion. It's the config you'll most often want to actually set, because realistic-person H3 outputs live or die on face stability.

    Like all six config nodes in this family, it doesn't produce a prompt on its own. It emits a typed 人物一致性配置 port that plugs into the Advanced Assembler's matching input; the assembler folds it (plus the other five categories) into a stateless English instruction for a remote LLM that writes the final MiniMax H3 prompt. Leave it unconnected and the assembler rolls a reproducible random identity config instead - which means your faces can drift run to run. Connect it if you care, and you probably do.

    The inputs that matter most for realistic subjects:

    • 身份锁定 - defaults to 严格 (strict), which instructs the LLM to preserve facial structure, hairstyle, skin tone, age feel, build, and clothing. This is the setting to keep at strict for anything with a real person in it.
    • 面部动作 - defaults to 微表情 (micro-expressions), and the tooltip spells out the trade: for realistic people, keep it at micro or small. Bigger facial motion is where identity starts sliding.
    • 头部动作 - head movement amplitude; subtler is steadier.
    • 面部朝向 and 视线行为 - keep the source direction and gaze, lock to camera, or allow small angle changes. "Locked to source" is the stability pick.
    • 眨眼 - no blink, one natural blink, or natural random blinks. Oddly important: no blinks at all reads as uncanny, too many reads as broken.
    • 嘴部动作 - the trap option. 说话 and 演唱 (speak/sing) add real facial-motion risk and require the model to animate a mouth convincingly. Default 保持原状 is the safe route.
    • 面部遮挡 - whether hands, hair, props, or effects may cover the face. Default 禁止 (forbidden), which keeps the face readable.

    Output: 人物一致性配置, a custom typed port that only connects to the assembler's identity slot. The whole pack is Chinese-labeled; the final prompt is always English.

    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.

    Where people get burned

    The classic mistake is flipping 嘴部动作 to speak or sing and then wondering why the face warps - talking is the highest-risk thing this category can ask for. If you want dialogue, budget for it: pair it with a strong 身份锁定 and keep 面部动作 small. The second stumble is leaving the whole category unconnected and then complaining the face changed between runs - that's the assembler's random fill doing its job. Connect the node and lock in your values.

    CategoryTlant Toolkit/MiniMax H3提示词/高级配置

    Inputs (9)

    NameTypeDefaultDescription
    身份锁定COMBO严格严格模式优先保持五官、发型、肤色、年龄感、体型和服装。
    面部动作COMBO微表情控制面部肌肉与情绪变化幅度;写实人物建议使用微表情或小。
    头部动作COMBO轻微控制点头、抬头和转头幅度;幅度越小通常越稳定。
    面部朝向COMBO保持原图控制人物脸相对镜头的角度变化。
    视线行为COMBO保持原图控制眼睛看向镜头、跟随镜头或短暂移开。
    表情变化COMBO保持原表情控制表情的发展方式,避免写实人物脸部突然变化。
    眨眼COMBO自然一次控制视频中的眨眼次数。
    嘴部动作COMBO保持原状说话或演唱会增加脸部变化和口型生成难度。
    面部遮挡COMBO禁止限制手、头发、道具或特效遮挡脸部。

    Outputs (1)

    NameTypeDescription
    人物一致性配置TLANT_MINIMAX_H3_IDENTITY_CONFIG