48.7 million views from an AI character with no visible eyes: how to design an AI character people remember

A TikTok account posting entirely AI-generated videos of a fake "bodybuilder" whose eyes you never see, and nobody knows who's behind it — yet the first 12 videos on the page combined got around 48.7 million views. This post summarizes and rewrites in Vietnamese the analysis of a creator who dissected that phenomenon and built his own AI character: why that character is memorable, the design process from reference sheets to video, costs, and why this wave might cool down fast.

This post is based on "$48K and 48.7M Views From an AI Bodybuilder Whose Eyes You Never See. So I Built My Own AI Character" by GHOST (@ghosstty_) on X (16/10/2026), rewritten in Vietnamese by OPCAIOS, condensed and with notes added for content creators in Vietnam. The numbers about the account mentioned are counted or cited by the original author; OPCAIOS has not independently verified them.

Derek Mercer: an AI character with no face

Derek Mercer is a bodybuilder character who never stepped foot in a gym — every video is AI-generated, and some captions even note it straight up as made with Higgsfield (an AI image and video generation platform). According to the author, the TikTok account has around 170,000 followers and 4.2 million likes, follows nobody, and the person behind it has never shown their face, name, or voice.

Before trusting numbers floating around, the author counted the views on the first 12 videos on the page himself: total around 48.7 million views, with the highest video around 20.4 million.

As for money, the author separated that out. Some posts on X say the creator of Derek released a token on the crypto platform pump.fun and collected around 48,000 USD in transaction fees; another account said 60,000 USD. The author does not treat that as income: token fees are more like a casino receipt than a salary, and vanish when the hype dies down. View counts are different — they sit public on the page, anyone can check them in two minutes.

Lesson 1: face is not required, body shape is

Most trend analyses on AI characters say the face is everything. But Derek's eyes are nearly always hidden by his bangs — the part that usually helps you recognize a hidden face — and he's still the most recognizable character on the author's feed.

Because everything that makes Derek is under his hair:

• Black emo haircut, long.
• Muscles so absurdly huge they become a joke.
• The same outfit in every video: gray shirt with blue print, studded belt, chain wallet, black cargo pants.

You recognize him from the shoulders down, from behind, at stamp size. Then the caption adds a twist: "Am I scary?", "I feel lonely…" — a sad teenager in the body of something that could flip a car, and viewers stick around to see which side wins.

The face might not be needed. Body shape is mandatory.

Lesson 2: describe your character in one sentence

Before writing the first prompt, the author set a rule for himself: if you can't describe the character to a friend in one sentence, the character isn't done. Derek passed the test in five words: emo hair, absurd muscles, gray shirt. That sentence is the product; everything else is decoration.

For his own character, the author picked four traits big enough to survive in every camera angle: a hairstyle readable from across the street, eyes a bit too surprised, front teeth slightly gapped when smiling, and a suit in a color that acts like a logo.

One more test from Derek's own page: on a phone, each video square in the grid is only about thumb-width wide, and Derek is still recognizable in all 12. So every character design had to be recognizable when shrunk down to grid size.

Process: three reference sheets, one character retained

The author builds a character using a character reference sheet: a single image on a smooth gray background, showing full body front view, side view, back view, plus two close-up face shots — so every video afterward uses one shared "original version."

1. First sheet — discard immediately. The character is strange, but intentionally so: hair like a hat placed on the head for a photo, a blue velvet tuxedo that looks like a wedding party staff member, a face tilted toward caricature. This kind of "strange" people scroll past, because it demands attention rather than earns it. The author keeps three things: round surprised eyes, sparse teeth, and a square hairstyle.
2. Second sheet — right face, wrong outfit. Rebuilt from scratch around the chosen features: black pillar hair standing tall almost level with the face, eyes exposing white around the pupils, sparse teeth with thin mustache and sideburns. Face is done. But the cream vest with black shirt and yellow tie "looks very nice" — and very nice is something no one remembers. Shrink it to a grid square, cream color fades into gray background, only the hair shows.
3. Third sheet — edit, don't rewrite. This is the step that changes everything.

Edit images instead of writing new prompts

A new prompt produces a new person. Even if you paste the old description exactly, the model still "rolls the dice" again and hands you… the character's cousin. But editing keeps the person you like and only touches the part you specify.

The author attaches the second sheet and requests: keep everything — face, eyes, teeth, sideburns, skin, body, hair, background, lighting, composition — only change the outfit, identical from every angle. New outfit: blue silk vest with woven pattern of vines and golden bees, pink silk shirt with big bow, gold chain and rings.

The character retained is named Cornelius. The reason for keeping him fits in one sentence: blue, pink, and gold is a combination no one wears by accident. Shrink him to a grid square and he's a blue block with a black hair pillar on top — recognized before seeing any facial feature clearly, even from behind (important because many movements borrow from viral videos with back shots).

The lesson wrapped in one line: design the face for close-ups, design the outfit for thumbnail images.

Give a character a "shop" before giving him a joke

Derek lifts weights — that's his whole world: a gym, a balcony, a body built for exactly one joke. The author wanted Cornelius to own something, because a business gives you hundreds of free content episodes. Cornelius becomes the owner of a luxury fur shop — not a sales clerk. That one word changes how he moves: a seller tries to convince you, an owner just shows you what's his and waits for you to understand.

First video script:

• Camera angle: phone placed on a stone counter, angled slightly up — exactly like a real shop owner filming product for social media. From this angle, the tall pillar hair under the ceiling fan makes him look twice the size of the room.
• Action: introduces each item straight into the lens with the pride of someone who hand-picked the goods.
• Setting: cream-colored and honey-colored coat racks; two staff members in tuxedos stand still in the back. No one reacts to the blue silk vest, because he's the boss — the world treats him as completely normal, and that's exactly what makes him strange.
• Final beat: he pushes a large white fur piece straight into the lens until it fills the frame, only his wide-open eyes peeking up from above — the only shot where the face nearly vanishes, and it's the shot people screenshot.

There's one more subtle reason: someone introducing products in front of a lens is already an advertising format. If a brand collaboration comes later, the product placement already exists in the show — he picks up the product exactly like he picked up the fur piece.

Where to Put the Reference Image: "reference", not "first frame"

The video was made using Seedance inside Higgsfield, with a full-body front-facing photo attached as a reference image, not as the first frame. It sounds technical, but the difference decides everything: use it as the opening frame, and the video opens with a studio photo on a gray wall and dies in hallucination half a second in; use it as a reference, and the model takes the face and outfit into a completely new room.

Three sentences in the video prompt carry most of the result:

• "The phone never moves" — so the model doesn't make up a cameraman.
• "The employee in the back stands completely still" — so the background doesn't steal attention.
• The push of the hair into the lens at the end — gives the video an ending. Most AI videos don't end, they just stop.

With dance videos, the author believes many videos in this wave take a dance that already went viral performed by a real person, then use motion transfer tools (for example Kling Motion Control) to make the character move identically. Body language borrowed from the real person, so it looks real.

Cost: little money, lots of time

The most detailed public breakdown the author found estimates 10 complete videos costing around 155 USD: about 102 USD for the video part (320 seconds, about 0.32 USD per second), about 3.35 USD for 50 images, the rest for other AI calls, editing and music. Counting redos, budget 200–300 USD.

The real cost is time: from the same source, it estimates 30–50 hours for the first 10 videos while learning, and after one evening making a reference sheet, the author found that number to be real.

This wave might be cooling down

Most articles about Derek leave this out: according to the author, Derek's first videos got around 2 million views each, while new videos struggle to break 100 thousand. The French character with the mustache who opened this whole wave is said to have lost about half its engagement after viewers figured out it was AI.

Novelty is the spark, not the fuel. A strange appearance buys the first million views and nearly nothing after. What survives the spark is a habit viewers want to see again: the same person, the same outfit, one new small situation each time. Cornelius's shop next week sells hats, then perfume, then a very expensive pair of gloves — a format that never runs out. Build a show, not a moment.

The author also raises two unsolved risks: dense patterns (tiny yellow bees) are the kind of detail video models struggle to keep stable, flickering when they move; and the formula is easy to copy — if Cornelius works, someone will make a character with a bun and blue vest suit in a week.

Mark "AI character" from the start

A decision the author made in advance: mark "AI character" in the account bio from the first post. Viewers click to learn how the character was made, and hiding it just makes comments angrier when they find out.

For content makers in Vietnam, this is also compliance: TikTok, YouTube and Meta all require labeling AI-generated content that looks real, and when the character starts taking ads, the ad content still follows ad rules like any real KOL — it's not that the character is virtual so it gets to oversell product benefits.

If you want to try: start with 10 videos and one question

The roadmap the author proposes for themselves:

1. One character, describable in one sentence, recognizable at grid-square size.
2. One reference sheet locked in, every change after that is image editing, not writing new prompts.
3. One repeatable setting (shop, counter, kitchen…) instead of a one-off joke.
4. Ten videos made by hand, with a budget you can afford to lose.
5. Not a second character yet, not automation yet, not tokens — until viewers ask for the next episode of the first character.

As the author writes, an AI agent can post a hundred videos a day, but can't fix a character nobody wants to see twice. After ten videos, you'll know if an outfit can become a character.

Source: GHOST (@ghosstty_) on X, 16/10/2026 — rewritten and condensed by OPCAIOS; account numbers and costs per original author.