Spellbrush.

It is a Tuesday in the Mission and someone at Spellbrush is, very seriously, arguing about eyes. Not eyes in general. The specific kind of eye that appears in a 90s shoujo manga panel - the one with the small triangle of highlight, the eyelash that flicks up at a particular angle, the iris that catches three colors instead of two. A model is being retrained. A character pose is being revised. Across the room someone is balancing turn order in a tactics RPG. This is what a generative AI company looks like when it is run by people who actually like the medium they are generating.

Spellbrush sits in a quiet pocket of San Francisco's AI scene, which is a strange thing to say about a studio whose co-built model has produced more anime images in the last two years than most human illustrators will draw in a career. The studio is small. Thirty people, give or take. It does not court the press. Its homepage is mostly four buttons.

And yet, click one of those buttons - the one marked niji・journey - and you arrive at a tool that quietly became the default for an entire visual subculture. Anime fans, illustrators, indie game developers, doujin authors, kids who are eleven and have a Discord account: niji・journey is now part of their muscle memory. The model was built in partnership with Midjourney and tuned, in Spellbrush's words, with about twenty-three billion parameters trained for one specific kind of beauty.

That one sentence does most of the work of explaining the studio. It is honest, it is small, and it is the opposite of how most AI companies talk. There is no manifesto. There is no claim that artists will be obsolete in eighteen months. There is the word "good," and the word "anime," and the word "game," and a quiet sense that these three things deserve to be in the same sentence.

The founders - Cory Li, Ruwen Liu, and Haitao Mao - met before the current AI cycle was a cycle. Li had been at MIT studying bioengineering, then at Benchling, the lab-software company. The pivot from cell cultures to chibi characters is the kind of jump that makes more sense once you understand the through-line: Li is the type of engineer who is also a fan. Spellbrush, the company, is what happens when fans are given a small seed check and the freedom to build the tools they wished existed.

That seed check, when it came, was modest. Y Combinator's W18 batch, about $130,000 in disclosed early funding, a few follow-on investors including Merus Capital, NewDo Venture, and Quest Venture Partners. Total funding to date, by public estimates, sits under three million dollars. In the era of nine-figure AI rounds, this is approximately a rounding error. Spellbrush appears to be fine with this.

What the studio chose to do with that money is build, not raise. First came Waifu Labs, a 2019-era generator that let anyone with a browser produce custom anime portraits. It went viral in the way things went viral before "AI generator" was a phrase your grandmother knew. Then came Cake Duel, a small playful game. Then Arrowmancer, an in-development anime mobile tactics title with a one-line pitch - "witches in space" - that should be hung in a museum. Then, working with Midjourney, niji・journey itself: the model that put Spellbrush on every anime artist's bookmarks bar.

The fact that all of this came from one studio, in one city, with one product page, says something about how Spellbrush thinks about leverage. It does not try to do everything. It tries to do anime, and to do it the way someone who loves anime would do it. That is rarer than it sounds.

The honest version of the company's pitch goes like this: general image models are good at almost everything and great at nothing. Ask one for an anime character and you get a face that looks like it was assembled by committee. The line is wrong. The eye is wrong. The chin is the wrong shape. A model purpose-built for anime, fed on the right data and tuned by people who can tell at a glance whether a piece of lineart "feels" right, produces something different. Not just technically correct, but visually fluent.

That fluency is what artists are willing to pay for. niji・journey runs on a subscription model, the same tiers Midjourney uses - ten dollars a month at the entry level, sixty at the top. A Discord-first product with an Android app on the side. A community in the hundreds of thousands. None of this is venture-scale by 2026 standards. All of it is profitable in a way that matters to a thirty-person studio that wants to keep making games.

That last part is the trick. Spellbrush is two things at once: an AI tools company, and a game studio. Most companies that try this end up doing one badly. Spellbrush appears to be running them as a single pipeline. The model produces art. The art trains the team's taste. The team's taste trains the model. Arrowmancer benefits from all of it. So does niji・journey. So does whatever comes next.

You can hire there, currently, as an LLM engineer, an AI researcher specializing in anime, a Unity engineer, an AI infrastructure engineer, or a frontend engineer working on the games. The job listings are unusual in their specificity. They are not asking for generalists. They are asking for people who can do one thing and care about a second thing - the difference between hiring an engineer and hiring a partner.

Back at the studio, on that Tuesday, the eye discussion ends with a small change to the model and someone going for coffee. Nobody high-fives. The anime keeps getting better. The game inches closer to ship. This is the thing about Spellbrush that is hardest to convey in a profile, which is that the work is mostly quiet. The company has no interest in being the loudest AI studio in San Francisco. It is, by a wide margin, one of the most interesting.

It is still Tuesday at Spellbrush. The eye conversation has resolved. Somewhere in the studio, a new character is rendering frame by frame. niji・journey, downstream of that conversation, is now drawing a slightly better version of itself. A player on a phone in Manila gets a sharper portrait. A tactics designer two desks over moves a witch into a slightly better position on a turn-order tree.

None of this is loud. None of it requires a press release. The model improves. The game inches forward. The studio stays at thirty. The work goes home with the people who made it.

This is the thing Spellbrush keeps quietly proving: that AI, done by people who actually love the thing the AI makes, looks less like disruption and more like craft. A long, patient sentence. A tool for artists, built by artists. An anime studio that happens to ship a model. A model studio that happens to ship anime games. Same building, same Tuesday, same eyes.