The AI companion industry had every tool it needed to build a realistic face, and much of its most visible part chose a drawing instead. The drawing is usually anime, and when it moves, it usually moves with Live2D, a Japanese technology that turns a single illustration into a character that blinks, breathes, and tilts its head. The choice is cheaper, safer, and better liked than a face that is almost real, and all three reasons come from the same place: a drawn face never has to be anyone.
What Live2D actually is
Live2D is an animation technique for bringing a flat illustration to life without drawing every frame and without building a 3D model. An artist paints a character in layers, eyes, lashes, strands of hair, the mouth, the collar, and a rigger then warps and rotates those layers so the figure can turn slightly, follow a gaze, open its mouth in time with speech, and shift expression. Simple models use around 50 layers; elaborate ones use hundreds.
The technology comes from a Tokyo company founded in 2006 by developer Tetsuya Nakashiro, which later renamed itself after the product. It spread through mobile games first, and then found its real audience: virtual YouTubers, the streamers who perform as animated characters. By the company’s own user data, 70 percent of people using its professional Cubism software are VTubers or make models for them.
That history is the key to the whole trend. By the time AI companions needed a face, an entire industry already knew how to make a drawn character feel present on screen, cheaply, on an ordinary phone.
Why the drawing wins
It is cheap and it runs anywhere. A Live2D character is one illustration and one rig. It needs no 3D pipeline, no motion capture, and no video generation, and it animates in real time on the phone already in the user’s pocket. For a small team building a companion app, that is the difference between a character with a face and a character with a profile picture. Real-time video of a realistic human face, by contrast, is expensive to render and expensive to get wrong.
It is liked more than half-realism. The best-known study on the question, by Rachel McDonnell and colleagues, rendered the same animated character in 11 styles, from pencil-sketch cartoon to near-photographic human, and asked people to rate them. The cartoon styles were rated the friendliest. The most realistic and the most abstract styles were both rated appealing, and the styles in the middle, too real to read as a drawing and not real enough to read as a person, were rated the least appealing of all. Motion made it worse: an unappealing character looked even less appealing once it moved. A stylized face sidesteps that trap by never promising to be human in the first place, which is also why the unease people feel with AI tends to come from somewhere other than the rendering, as the research on the uncanny valley in conversation shows.
The audience already exists. Anime style arrives with a fandom that knows how to relate to drawn characters, spend on them, and follow them for years. The artists, the riggers, the model marketplaces, the conventions of expression and voice were all built by the VTuber economy. An AI companion wearing that style inherits a visual language its users already speak.
Where Live2D anime companions stand in 2026
The clearest proof of the format comes from the human side of it. Cover Corporation, the company behind the VTuber agency hololive, reported net sales of about 49.3 billion yen, more than 300 million dollars, for its fiscal year ending March 2026, up 13.7 percent. Most of that came from merchandise, live events, and licensing built around drawn characters. The format carries a real business.
The clearest proof that it carries an AI comes from a single channel. Neuro-sama, an AI VTuber built by an independent developer, debuted on Twitch in late 2022 wearing a free sample Live2D model, then two custom Live2D models over the next two years. In early 2026 she passed 160,000 active subscriptions and became the most-subscribed channel on Twitch, ahead of every human streamer, and set Twitch’s all-time hype train record. For that stretch, a language model with a drawn face had more paying subscribers on Twitch than any person on it.
Below those two, the format is spreading from the bottom up. Open-LLM-VTuber, an open-source project for talking hands-free to any language model through an animated Live2D character, has passed 13,500 stars on GitHub. The app stores now carry a crop of small companion apps with Live2D in their names. The ones that surface first lean romantic, and at least one of the most visible sells gems and memberships, with users complaining in reviews about energy limits that stop the conversation until they pay or watch an ad.
At the top of the market, the money moved one step further. Love and Deepspace, the largest revenue product in companion romance, is an anime-style game rendered in 3D, and its economics are told in the billion-dollar AI boyfriend. xAI’s Ani, one of the most publicized companion launches of 2025, was an anime character built as a 3D animated avatar. Neuro-sama herself moved to a 3D model late in 2025. The trend is consistent: Live2D is where anime companions start, because it is cheap and it works; 3D is where the ones with money go. What almost none of them does is leave the anime style behind.
The part the VTubers kept
Look again at the business that proved the format. A hololive talent is a drawing on the screen and a real person at the microphone. The fans know it. The drawn face is worn by someone with a history, a temperament, and a life that goes on off stream, and the bond fans form is with that someone, seen through the drawing. The stylization never had to carry the relationship alone. It only had to carry the person.
Most anime companions kept the drawing and dropped the person. The face is still expressive, the voice is still warm, the Live2D rig still blinks on cue, and behind it is a model with no life that goes on anywhere. That is exactly why the style suits them. A drawing makes no claim about who is behind it, so no one can catch it failing the claim. It is the safest possible face for software that is not anyone: friendly by design, protected from the uncanny, and never asked the question a realistic face invites at once, which is whether there is actually someone here.
That question does not go away because a company chose not to invite it. It is the question at the center of what real Human AI is, where the answer is built on a real person who exists in real life rather than on a rig. The anime companion answers it by drawing a face that does not need an answer.
Every hololive model on a screen has a person breathing behind it. Most anime companions have a rig that breathes on a timer.
Sources: Live2D Inc. (company history and Cubism user data, as compiled by Wikipedia’s Live2D entry). McDonnell, Breidt, and Bülthoff, “Render Me Real? Investigating the Effect of Render Style on the Perception of Animated Virtual Humans,” ACM Transactions on Graphics (SIGGRAPH 2012). Cover Corporation, financial results for the fiscal year ended March 2026 (Tokyo Stock Exchange disclosure, May 2026). Tubefilter (January 2026, Neuro-sama subscription record); Wikipedia, Neuro-sama (model history, citing Kotaku Australia, Bloomberg, GameSpot). Open-LLM-VTuber, GitHub repository. Apple App Store listings (Live2D companion apps, in-app purchases). AppMagic estimates via PocketGamer.biz (2025, Love and Deepspace). TechCrunch and TestingCatalog via Techmeme (July 2025, Grok companions).
