Five characters. Six languages.
Drawn on your device.
New characters
Individual refinement
Cross-language testing
One documented Polish QA run · 15 Sept
Preparing film
Checking the recording package…
Live avatars · prerecorded Azure speech
Rendered on your device. Not streamed video.
Space to pause · Esc to exit
WHAT IT SOLVES
WITHOUT STILLTALK
You need a voice and a face that moves with it. With a cloud-rendered avatar, you can choose a separate voice provider, or get voice and video from the same service.
Either way, the avatar still needs to be rendered in the cloud and streamed to your users.
That means video bandwidth, another source of delay, and the cost of generating the moving image—not just the voice.
WITH STILLTALK
Instead of streaming video, we bring the character to life on your user’s device. Speech arrives with compact movement data. The device draws the animation in sync with the voice.
And it’s more than lip sync. Subtle eye, head, hand and leg movements bring the whole character to life.
Load the character’s graphics first. After that, there’s no continuous video stream to buffer and no external video-rendering service charging by the minute.
No video stream. No cloud video renderer.
Built for lower running costs. API pricing comes next.
Want to build with it? Get product updatesThe demo is live. Get notified about the API and custom characters.
BEHIND THE DEMO
Not one prompt. A working partnership,
built through experiments and close review.
01 — THE FIRST EXPERIMENTS
We began with a more traditional 3D renderer. We also tried different models and approaches. For this project, the early results didn’t meet the quality bar.
Working with Astra, we kept iterating—towards a photo that could feel present, not just move its mouth.
02 — SEPTEMBER 13–14
Clara came first. The first versions of Julian, Nia, Pistak and Ren followed in one intense day, September 13–14. They’ve been refined since.
The images were generated with an image model under Astra’s direction. Each character was then measured, rigged and calibrated individually.
03 — THE LONG NIGHTS
A blink sitting too high. A hand pulling on a sleeve. A mouth shape that worked in one language, but not another.
Long sessions meant checking individual frames, testing real speech and coming back to what still looked wrong. The founder set the quality bar; Astra implemented and tested the changes.
04 — THE SCALE OF THE WORK
One overnight calibration alone ran from 9:19 p.m. to 6:21 a.m. Five characters, six languages, and more than two million automated geometry checks.
The numbers describe the work. Close visual review decided what needed another pass.
Tokens: at least 1.8 billion input and output tokens logged in this development thread, September 4–17. This includes repeated and cached context—not 1.8 billion newly written tokens, a bill, or total account usage.
Pixels: 305,510,400 pixel comparisons across 1,989 frames in three release checks. Animation: 51,748,488 values compared. The September 14–15 calibration recorded 447,157 rendered frames and 2,108,019 geometry checks.
These are documented automated workloads, not a claim that every frame was personally reviewed. Rendered test frames are not AI image generations.
The day we bought
stilltalk.ai.
05 — A NEW DEADLINE
We discovered the Astra challenge on Saturday, September 12. The engine was already working. The next day we started building new characters. On Tuesday, September 15, we bought stilltalk.ai.
The next push was getting it into people’s hands: a public demo, live speech and repeated tests on phones—not just on a development laptop.
Build.
Watch closely.
Refine.
06 — STILL MOVING FORWARD
The characters share an engine, but each has its own calibration. We’re refining articulation across languages and the small movements that make each character feel natural.
Our ambition is simple: precise lip-sync and natural expression, without making visual quality the price of a lighter experience.
GPT-6 Astra designed and implemented the rendering engine, articulation, calibration and tests. The founder guided the product and reviewed the results throughout. That collaboration continues.
07 — WHAT COMES NEXT
A talking character should be something you can build into your product—not something your product has to work around.
We’re working toward a rich character library and a beta for creating avatars from your own photos. Carefully calibrated lip-sync. Natural movement. Lightweight delivery.
The goal is to make expressive avatars accessible to more products, with less bandwidth and lower running costs than cloud-rendered video.
01 / 07
YOUR TURN
Pick a character. Choose a line. Press Play.
Rendered live, right here on your device.
Choose a character
Rendered on your device. Not streamed video.
Loading recordings…
A photo that talks back.
WHAT’S NEXT
Five characters, six languages, your own sentences. That’s live today—try it yourself.
A commercial API and a wider library of avatars and animations are in development, alongside a beta for creating a character from your own photo. Leave your email and we’ll tell you when they’re ready.
stilltalk.ai · Try it yourself