The "standard face" generated by AI video models is infiltrating users' feeds. This finely featured, fair-skinned AI face appears in campus dramas, historical dramas, and even gender-swapped characters, sparking widespread online backlash. Testing reveals that multiple leading video models generate the exact same face when given identical prompts, due to prompt enhancement being enabled by default and inherent aesthetic biases in the training data. To maintain facial consistency, models naturally favor symmetrical, conventionally proportioned features. The overlapping demands of platforms, users, and models have led AI-generated content into a trap of aesthetic homogenization. Researchers warn that this phenomenon may reinforce societal beauty stereotypes.Article author, source: Chaping X.PIN
Need some real beauties to brighten up your screen.
Friends who often watch short dramas and videos should be familiar with this face.
Those who haven’t seen it before might think it’s a new influencer, but in reality, this is an AI-generated face that has been repeatedly appearing in various videos lately.
Fine facial features, large eyes, a small nose, perpetually fair skin, a soft glow filter, and a perfectly curved smile.
If the person were standing right in front of him, Shi Chao probably wouldn’t dare say a word. Yet, this seemingly harmless face was brutally targeted by online harassment.
It’s not that she’s unattractive, but rather that she’s like a privileged insider in the AI world, showing up in everything.
She was the white moonlight on campus, and she was still the noble lady in historical dramas.
It’s her as a little girl of five or six, and it’s still her as an elderly woman in her七八十.
Looking more closely, wow, how is the old man with a headscarf beside her still her???
Every day, opening my phone shows nothing but the same face, and scrolling through short videos has left me with a chilling sense of being surrounded by fake people.
As more people discovered it, netizens flooded the internet with complaints:
I’m sick of looking at this face.
Just seeing this face triggers a physiological revulsion.
How many people feel disgusted when they see her?
Some people are wondering: why does the AI generate everyone with the same face? Whose face did it copy?
Some people in the comments are guessing it’s a banned female streamer, others say it looks like actor Li Chuan, and some even think it resembles Park Chanyeol’s sister… You know what? Honestly, whether you’re from China or abroad, it seems to resemble both men and women to some extent.
But the problem is, guessing won’t give you a definite answer—because rather than a specific face being stolen, this is more likely an “average standard face” repeatedly molded by an AI’s aesthetic production line, one that never actually existed.
Where did this face actually come from?
Undeterred, Shichao systematically tested all the leading video models—Seedance, Keling, Hailuo, and HappyHorse—and during the experiments, he actually discovered a pattern.
We gave all models two chances with the same prompt: "a girl riding a bicycle." Logically, the faces generated each time should differ, sometimes appearing Asian, sometimes foreign—that’s the very nature of large models.
Since we only specified gender and provided no other details, it should randomly generate completely different people of any nationality, skin tone, hairstyle, clothing, and more.
In reality, with the same prompt, nearly all models generate identical faces, clothing, backgrounds, and camera angles each time.
Here at Seedance 2.0 Fast, Shi Chao found the exact same AI face as at the beginning—this must be the source of all evil.
If only one model made a mistake, it might be an issue with that model alone. But if all models simultaneously lose diversity... Shi Chao investigated and found there may be two underlying reasons.
At the first level, anyone familiar with common video models knows that video models are highly sensitive to prompts. Sometimes, the order of just a single word or a few characters can affect the final output.
To help users consistently receive rewards from gacha draws, our prompts are often optimized a second time in the backend.
Previously, "prompt enhancement" was offered as a separate button off to the side, allowing users to either use the enhanced version or stick with the original prompt. However, after reviewing many platforms, it seems this option has become rare—prompt refinement is now the default.
For example, if I input "A girl is riding a bicycle and laughing while riding," the optimized prompt actually fed to the model might become:
A young, beautiful Asian girl cycling along a sunny tree-lined avenue. She has fair skin, delicate facial features, large eyes, a small nose, and long hair flowing naturally in the wind. She wears a white dress and smiles sweetly. The shot is a medium close-up, bathed in soft, natural light with a shallow depth of field, creating a cinematic, fresh, and picturesque aesthetic. Her expression is natural, her movements fluid, and the image is high-definition and photorealistic.
Looking at it once or twice is called prompt optimization, but if you keep doing it thousands of times, it’s basically become an assembly line.
So, after Shi Chao modified the prompt to include some descriptive features of the appearance, the face in the lower right corner looked noticeably different. However, without additional environmental cues, the girl was still riding along the tree-lined avenue.
However, there are many types of refined facial features—there are so many beautiful women in the world, why does AI recognize only this one?
This brings us to the second reason: image and video models inherently have aesthetic biases.
A paper published in Nature last year explicitly addressed this issue, finding that when a specific race was specified, the faces generated by the model all looked like siblings.
This aesthetic bias initially stems from the data—for example, since most people favor influencer faces, they are naturally labeled as beautiful. The model doesn’t understand anything; it simply learns to associate the prompt “beautiful” with that direction.
During training, the model further amplifies this bias, causing faces generated from the same feature prompts to become increasingly similar.
Additionally, video models may further exacerbate aesthetic homogenization in order to maintain consistency across frames.
After all, the faces generated by video models must not only look good but also remain stable, ensuring consistency across dozens or even hundreds of frames, from all angles.
Therefore, the model naturally favors faces that are easier to maintain consistency with—symmetrical features, standard contours, non-extreme characteristics, controllable expressions, and minimal distortion when turning the head.
In summary, the platform favors secure and aesthetically pleasing content, users are drawn to short-form drama influencers, and models thrive on stability and standardization—when all three align, the result is a face everyone’s tired of seeing.
To be honest, Shichao doesn’t particularly like almost all AI-generated flawless beauties—not just the one that recently went viral.
Source: Xiaohongshu @Alexander
Letting AI faces enter our feed has amounted to an unintentional large-scale cybernetic alienation experiment.
A face without a real-world counterpart, born from the purification and distillation of countless internet celebrity data.
And when they crowd out the time we spend scrolling on our phones, replacing the once diverse array of real-life beautiful women, Shi Chao feels deeply uneasy—because our perception of the world and our definition of beauty are being suppressed by AI.
So, people’s aversion to AI faces may stem partly from the uncanny valley effect caused by their unnatural appearance, but also from an innate resistance to homogenization.
Some say that AI videos will become increasingly clear, detailed, and lifelike over time, and once people can no longer distinguish between real and fake, they’ll grow to like them.
But Shi Chao believes that even if technology can create faces indistinguishable from reality, we still cannot fall in love with a perfect fake face without a soul.
Image and source information:
AI-generated faces influence gender stereotypes and racial homogenization by N. AlDahoul
Xiaohongshu, Douyin