WITH THE EXPLOSION OF VIDEO GENERATION TOOLS SUCH AS DREAM AI, COVEN AND CONCH, WE FOUND A PAINFUL POINT: IT'S EASY TO CREATE A NICE FACE, BUT IT'S HARD TO MAKE IT "NATURAL" IN FIVE, 10 SECONDS OF VIDEO。
DOES THIS HAPPEN ALL THE TIME: DO YOU WANT A "SMUGGLE" VIDEO, AND THE RESULT IS THAT THE PERSON CREATED BY AI GETS BIGGER AND BIGGER, UNTIL HIS CHIN IS DISLOCATED, OR HIS FACE MUSCLES MUTILATE LIKE LIQUIDS
This is itAI VideoThe resulting bottom-level logical defect: it does not understand “restraint”。
In the video stream, excessive emotional commands (e.g. “Angry”, “Laughing”) are tried and strengthened by each frame of AI, resulting in an “emotional overload”, which collapses。
Micro-expressions are not just for real, but for “stable”。
TODAY, WE DON'T TALK ABOUT PICTURES, WE TALK ABOUT VIDEOS. WE'RE USING THREE CORE METHODS TO TEACH YOU HOW TO CONTROL THE "TRANSFORMATION" OF THE VIDEO MODEL WITH A HINT, SO THAT YOUR AI CHARACTER HAS A NUANCED ACTURE AND NEVER LOSES FACE。
Method I: Control of emotional intensity - “restraint” for “stability”
1.1 "Overload effects" of video generation
IN TODAY'S TOP TOOLS (E.G. DREAMAI, CLIN) WE RARELY SEE FACES BROKEN. BUT HAS ANYONE NOTICED THAT THE SMILE THAT AI PRODUCES, WHILE PERFECT, ALWAYS MAKES PEOPLE FEEL "FAKE," "GREASY"
When you write:
A girl laughing loudly
NOW THE AI MODEL PRODUCES A PERFECT SMILE, BUT THE PROBLEM IS THAT THE SMILE WILL REMAIN AT THE STRENGTH OF 100% FROM SECOND TO FIFTH。
WHEN A REAL HUMAN LAUGHS, THE FACIAL MUSCLES WILL HAVE A SLIGHT TREMOR, THE THORAX WILL RISE AND FALL, AND THE SMILE WILL FALL. BUT THE VIDEO GENERATED BY AI TENDS TO BE LIKE A “DYNAMIC MASK” — THE RADIANS OF THE MOUTH ANGLE, THE BENDS OF THE EYE BEING PRECISELY LOCKED TO DEATH. THIS LACK OF BREATHING) IS PERFECT, AND THAT'S WHY THE AUDIENCE RECOGNIZED "THIS IS AN AI VIDEO."。
CORE CONCLUSION: AI'S DEFAULT “MAXIMUM VALUE” DOES NOT BRING ABOUT A COLLAPSE, BUT RATHER A CHEAP “PLAYING SENSE”。
1.2 Demersal strikes: “Unfinished” for “real sense”
WE CAN'T LET A.I. DO THAT. WE MUST INTRODUCE UNCERTAINTY AND A SENSE OF RELAXATION BY ARTIFICIALLY “NOISE” IN THE HINTS。
Miswriting (caused greasy):
Beautified women are happy at the camera.
(Results: Looks like a model with a stiff, fake smile for advertising. I'm not sure
Correct (micro-expression path):
A former woman, could play on the lips, soft gaze, facile muscles relaxed, not posing.
copy
"A relaxed woman with a light smile on her lips, a soft look, a loose face, no tweaking"
FIELD DEMONSTRATION (DREAM AI/ CLIN):
Group I: video alerts for negative teaching materials
DESIGN LOGIC: THIS IS A HINT THAT 901 TP3T WOULD WRITE. YOU JUST TOLD AI, "SHE'S HAPPY," "SHE'S HAVING COFFEE." CONSEQUENCES: AI CAN'T FIND A FOCUS ON ACTION, BUT IT CAN ONLY MAKE PEOPLE LAUGH AT THE CAMERA, OR SHAKE IN THE AIR WITH A CUP, COMPLETELY WITHOUT LIFE。
Prompt A
The woman is drining Coffee and smiling happy
The problem is "state word": Smiling happy is a conclusion。
AI Reaction: In order to meet “Happy” and “Drinking”, AI usually keeps people in a cup-side position, with a stiff smile on their faces, neither drink nor drop。
Group II: video alerts for positive teaching materials
Design logic: We don't write "happiness," we write "drinking." Consequences: Only by describing physical actions can facial muscles be naturally carried out in a manner consistent with the movement. You'll see lips sprung to blow and eyelids plunging to feel the scent。
Prompt B
Then she does not drink yet.
The core is "microactions": Smell the aroma - >Eyes close - >Blows on stage (blowing)。
AI REACTION: AI WAS FORCED TO CALCULATE THE OSCILLATION OF THE NOSTRILS WHEN THEY SMELLED, THE SHAPE OF THE LIPS WHEN THEY BLEW. THESE PHYSICAL MICRO-ACTIONS, WHICH DIRECTLY ELIMINATE THE "FALSE SMILE," MAKE THE AUDIENCE FEEL LIKE SHE'S A LIVING PERSON。
A template for the reduction of emotional strength
Application of scenery: prevent greasy, stiff, high-profile faces. Core logic: replace the word “mixed” with the word “micro-modified”。
Chinese Filling Formula (Direct Copy):
[Personal description], [emotional] on the face with [slight words], mouth horns [mob movement], eyes [eye state], face muscles relaxed, no slapping, breath, real nature。
Fill in an empty vocabulary (reference):
[whispering words] faint, subtle, barely invisible, subpressed.
[felts]: smile (frowning), freown (frowning), smirk (smirk laughing), worry.
[ muscular details].
[easy] relaxed, breathing naturally, cdid.
METHOD II: A CAUSE FOR MICRO-EMOTICS — ELIMINATION OF "AI" WITH "PHYSICAL INERTIA"
2.1 Why do you have a video character like a "wire puppet"
Many people produce videos, people with faces, but feel like they're being held by a wire. Like you wrote:
Shey woman.
♪ The shy woman ♪
In the video, the woman would stare at the camera, and then blush, twist. This is terrible, because the face of a real human being is a “by-product” of a physical movement, not an independent one。
IN VIDEO PRODUCTION, IF WE ONLY WRITE AN EXPRESSION, AI FAKES FACIAL MOTION; BUT IF WE WRITE AN ACTION, AI PRODUCES AN EXPRESSION BASED ON HUMAN MOTORS。
2.2 动作驱动法
公式：身体微动 + 视线转移 = 真实微表情
案例：表现“害羞”
只写表情：
The woman is shy and looking directly at the camera. She smiles shyly at the viewer. She maintains eye contact the whole time. Static head, just facial expression changing.
(中文大意): 这个女人很害羞，直视着镜头。她对着观众羞涩地笑。她全程保持眼神交流。头部静止，只有面部表情在变。
原理： 虽然写了 shy，但因为保留了 looking at camera（看着镜头）且没有写其他动作，AI 会把“害羞”理解为一种静态的扭捏，人物会死盯着观众挤表情，非常不自然。
加入动作因果：
She feels shy. She immediately lowers her head to avoid eye contact. Her eyes look down and dart to the side nervously. She tucks her chin in and bites her lower lip gently. She cannot look at the camera.
(中文大意): 她感到害羞。她立刻低头以避开视线。她的眼睛向下看，紧张地看向旁边。她收下巴，轻咬下唇。她不敢看镜头。
原理：
我们加上了 shy（害羞），但紧接着用 lowers her head（低头）和 avoids eye contact（避开视线）来承接这个情绪。 这样写的好处： AI 知道她“为什么”低头（因为害羞），生成的低头动作会更含蓄、更柔和，而不是像机械一样低头。
适用场景：防止“提线木偶感”，让人物不再傻盯着镜头，拥有肢体语言。 核心逻辑： 动作在前，表情在后。
Chinese Filling Formula (Direct Copy):
视频开始时，人物处于[起始状态：平静/发呆]。然后，[发生了触发动作/变化]，眼神[产生了反应]。最后，脸上慢慢浮现出[结束状态：情绪]。
Fill in an empty vocabulary (reference):
[手部/头部动作]: scratching back of head (挠后脑勺), tucking hair behind ear (挽头发), rubbing eyes (揉眼睛), looking down at phone (低头看手机).
[视线移动]: avoiding eye contact (避开视线), eyes darting sideways (眼神游移), looking down shyly (害羞地低头).
[引发的表情]: nervous smile (紧张的笑), relieved look (释然的表情), confused frown (困惑的皱眉).
方法三：给情绪加时间顺序——利用视频模型的“时序思维”
3.1 视频是时间的艺术
这是今天课程最硬核的部分。 方法三是视频生成区别于图片生成的根本逻辑。
图片是静态的，视频是流动的。如果你给视频AI一个静态描述（如“他很伤心”），AI不知道这个“伤心”在第1秒和第5秒有什么区别，于是它只能死循环。
微表情的本质是Change（变化）。 我们需要在Prompt中写入时间轴。
3.2 动态演变写法
目前的顶尖工具都能极好地理解“过程描述词”。
公式：Start (起始状态) -> Transition (变化动作) -> End (最终微表情)
剧情设计： 一个“如释重负”的电影镜头。从严肃的沉思，到深呼吸，最后因为想通了某事而露出释然的微笑。
Prompt:
The video starts with the man maintaining a serious, stoic expression, gazing into the distance. Then, he closes his eyes slowly and takes a deep visible breath (shoulders dropping). Finally, as he opens his eyes again, a faint, relieved smile slowly forms on his lips. Subtle movement of hair in the wind.
第一阶段（Start）： maintaining a serious… expression —— 锚定起始帧的状态，告诉 AI 先别乱动。
第二阶段（Transition）： closes eyes… takes a deep visible breath —— 这是核心！ 用“闭眼”和“深呼吸”这个物理动作，切断了严肃的情绪，为转变做铺垫。
第三阶段（End）： faint, relieved smile slowly forms —— 只有经过了前面的深呼吸，这个最后的微笑才显得有血有肉，而不是像假笑。
适用场景： 生成有剧情感的视频片段，防止画面静止不动。
核心逻辑： 起始状态 -> 触发动作 -> 最终结果。
中文填空公式：
视频开始时，人物处于[起始状态：平静/发呆]。然后，[发生了触发动作/变化]，眼神[产生了反应]。最后，脸上慢慢浮现出[结束状态：情绪]。
Fill in an empty vocabulary (reference):
[状态A]: neutral expression (平静表情), sleeping face (睡脸), bored look (无聊).
[触发动作]: eyebrows suddenly raise (眉毛突然上挑), pupils dilate (瞳孔放大), takes a deep breath (深呼吸).
[状态B]: warm smile (温暖微笑), shocked expression (震惊), tear falling (落下眼泪).
课程总结：视频Prompt的“减法”哲学
同学们，做AI视频，“少即是多，动即是稳”。
控制强度： 不要让AI全力以赴，限制它的发挥幅度，防止面部崩坏。
动作归因： 用挠头、低头、眼神躲避等物理动作，去“带出”表情，而不是“摆出”表情。
时间顺序： 描述情绪变化的过程，而不是结果。
现在的视频工具（即梦、可灵、海螺）都已经非常强大，缺的不是算力，而是懂得如何用文字引导它们“收着演”的导演。
希望大家课后去试着生成一个“不仅没有大表情，甚至还在努力掩饰表情”的视频片段。如果你能做到这一点，你就掌握了AI视频微表情的最高心法。