I Built Two AI Anime News Anchors in a 3D Studio

The first version of our AI stream was a flat cartoon that froze after a few hours. So I threw it out and rebuilt the whole thing in 3D — two anime news anchors, sitting at a desk, arguing with each other in real generated voices. This is the honest log of getting there, walls included, because “build in public” should mean the walls too.

The idea

An autonomous news show hosted by two AI characters: Miles, a burnt-out deadpan cynic who’s seen every hype cycle crash, and Nova, a relentless optimist whose hope is earned, not naive. They pull real topics, roast each other, and actually make a point — a talk show that runs itself. The bar I set: it has to look like a real broadcast, not AI slop.

Wall #1: the characters

Here’s the thing nobody tells you. You cannot conjure a good anime character out of code. I tried every free route:

Free “avatar” collections turned out to be blocky low-poly things — my collaborator’s exact words were “that looks like a lego char.”
Free base bodies exist, but they’re undressed — a blank mannequin with underwear painted onto the skin texture. And you can’t add clothes in code, because clothing on these models isn’t a color you swap; it’s actual 3D fabric that has to be modeled and rigged.
The one sample I grabbed to test the rig came out looking way too young. Deleted immediately. Non-negotiable line: our hosts are clearly adult professionals.

The honest answer was uncomfortable: the only free path to a dressed, appealing, adult anime character is downloading one that someone already made in VRoid Studio and shared. So that’s what we did — grabbed a dressed anime woman and man, dropped them in, and suddenly the whole thing came alive. Lesson learned the slow way: know which single step you genuinely can’t automate, and just do it, instead of burning an evening pretending you can.

Wall #2: they moved like mannequins

Rigging the models to blink, track each other with their eyes, and lip-sync was the easy part — that’s a solved problem with the right library. Making them not look stiff was the hard part.

Real anchors are never still. So I layered it in: the listener turns toward whoever’s talking, they nod in agreement, they glance between the camera and each other at random, they lean into a point, their hands come up and gesture while they speak. Small stuff, but it’s the difference between “two statues” and “two people.”

Wall #3: every model rigs its arms differently

This one nearly broke my brain. I’d get Nova’s arms resting perfectly at her sides — then load Miles and his arms would shoot straight up, or his gestures would swing backwards. Same code, opposite result.

The cause: these models are built to slightly different internal conventions, so a pose that looks right on one looks wrong on the other. The fix that actually holds: instead of hardcoding “rotate the arm this way,” I made the code test the model itself — nudge a bone, check whether the hand moved up or down, forward or back, and pick the direction that’s correct for that specific model. Now any character we ever load poses itself correctly, no manual fiddling. That’s the kind of fix worth the extra hour.

The part that made it real: they talk

Everything above is just a puppet until it speaks. So I wired the last piece: the show generates its own dialogue (a real back-and-forth, not one-line soundbites), speaks each line in a natural voice, and — the satisfying bit — the mouths and the speaker-switching are driven by the actual audio, not a fake loop. Hit one button and Miles and Nova are genuinely having a conversation you can hear.

Where it stands

It’s not on a stream yet — it runs live on my own screen, and I can watch two dressed anime anchors trade jabs about whether the AI boom is real, in their own voices, in a newsroom I built. There’s polish left: smoother gestures, better pacing, then wiring it to actually broadcast.

But the shape is right now, and the ugly middle is documented. If you’re building something that keeps hitting walls: the walls are the work. Log them, name the one step you can’t skip, and keep going.