01 · Overview
A full-size model that keeps up with a conversation
The most convincing digital humans have always been too slow to talk to. Maya-2 is our answer: a full-body human model, designed in-house at Thrifty AI, that starts streaming 0.90 seconds after it hears you and keeps going at 32 frames per second.
Many real-time avatar systems get their speed by simplifying the picture, and the result looks like it. We set out to keep the realism and make it fast, and built the model and the system that runs it together.
02 · Real time
Keeping up with a conversation
People judge a conversation by its pauses. If a reply takes too long to start, the exchange stops feeling live, however good the face looks. Maya-2 begins answering in 0.90 seconds and then streams continuously, so it can be interrupted, pick up again and hold a natural back-and-forth.
A live human has to look as good in the last minute as in the first.
A support shift, a lesson or a live event runs for hours, not seconds. Maya-2 is built for that length of conversation, not for a ten-second demo clip.
03 · Range
Any face, any style
One reference image is enough. Maya-2 animates photoreal people, illustrated characters and anime with the same lip-sync and the same natural upper-body motion, so your AI human can look exactly like your brand.
Stylised characters
Photoreal, full gesture
04 · Architecture
Architecture
Maya-2 is a 14B-parameter diffusion transformer that generates video one chunk at a time. Speech features, identity features and a few frames of recent motion condition every chunk. Inside a chunk every frame can see every other frame, so gestures stay coherent, and chunks are chained so the stream never has to stop.
Keeping that full-size model and still answering in under a second took three things working together: how the model attends, how it is distilled, and how it is served.
In
Transformer block × N · 14B
- Self-attention, bidirectional inside the chunk
- Reference cross-attention
- Audio cross-attention
- Feed-forward
Out
33-frame chunk28 new frames, plus 5 motion frames carried into the next chunk
Bidirectional inside every chunk
Most streaming video models only look backwards, which is where drift and flicker begin. Maya-2 keeps full attention within each chunk, so motion is planned with near-future context and long streams hold together.
Self-correcting distillation
Sampling is compressed to 4 steps with no guidance pass. The student is then trained to generate several chunks in a row from its own outputs, so it learns to repair the small errors it creates.
Full-stack acceleration
Attention, decoding and the whole execution graph were optimised in turn, so quality survives the trip to sub-second latency on a single server node.
05 · Systems
Real-time inference
Fast training alone doesn't make a 14B model real time. Maya-2 runs on a purpose-built inference stack, and every stage of the loop was optimised in turn.
- Audio processing 33 ms
- 4-step denoising 616 ms
- Frame decoding 187 ms
- Motion encoding 14 ms
- Other 26 ms
Systems work
~5×
Hybrid sequence parallelism
Attention is split across eight GPUs, cutting a single denoising step about five-fold.
~5×
Parallel video decoding
Once denoising is fast, decoding becomes the bottleneck. Splitting it across GPUs removes it.
−20%
Fused attention kernels
Overlapping data movement with compute trims attention latency by a fifth.
154 ms
Whole-graph compilation
Fusing the pipeline end to end brings one denoising step on 8 GPUs from 193 ms down to 154 ms.
Latency by GPU count (ms)
| GPUs | Video encode | Video decode | Denoise step | Denoise step, compiled |
|---|---|---|---|---|
| 1 | 97 | 988 | 1,070 | 800 |
| 2 | 69 | 690 | 620 | 490 |
| 4 | 39 | 350 | 313 | 261 |
| 8 | 21 | 192 | 193 | 154 |
06 · Evaluation
Evaluation
We compare Maya-2 against six leading audio-driven avatar models on short clips and on long streams. Maya-2 leads on aesthetics, image quality and lip-sync on short clips, and is the fastest model tested by a wide margin.
0.90 s
To first frame
From the first word of audio to the first frame of video, so a reply starts while the person is still listening.
32 FPS
Live video
Above the 25 frames per second that live video needs, with headroom for busy servers.
Full body
Whole-person motion
Upper-body gesture and posture come from the same model as the lip-sync, not from a separate animation layer.
No resets
Long conversations
Streams run for as long as the conversation does, without cuts, restarts or a change of face.
- Maya-232
- Fastest other real-time model~21
- Leading offline full-body models~5
- Typical diffusion avatar models<1
25 FPS real-time threshold
Quality on short clips
| Metric | Maya-2 | Best real-time peer | Best offline peer |
|---|---|---|---|
| Aesthetics score | 3.51 | 3.10 | 3.45 |
| Image quality | 4.79 | 4.37 | 4.70 |
| Lip-sync confidence | 1.47 | 1.04 | 1.32 |
| Lip-sync distance | 11.56 | 12.10 | 11.85 |
| Motion smoothness | 99.61 | 99.75 | 99.60 |
| Throughput (FPS) | 32.00 | 21.80 | 5.10 |
Higher is better, except lip-sync distance. Best result per metric across the models in each group, so no single peer achieves every figure. On streams longer than five minutes Maya-2 keeps the best lip-sync confidence tested (1.61 against 0.96 real-time and 1.51 offline). Peers that repaint only the face score higher on pixel consistency; Maya-2 generates full-body motion, which adds natural variation, and still holds 99.22 subject consistency.
Evaluation set: 100 clips under 10 seconds and 20 streams longer than 5 minutes; tested out to 1,000 seconds of continuous generation. Peers are anonymised groups of six leading audio-driven avatar models.
07 · Production
Built for production
A fast model is only useful if it is fast in production. Maya-2 ships with the Maya runtime, a serving stack we developed alongside the model, so the same sub-second start-up holds whether it runs in our cloud, in your own cloud account or on servers in your building.
- Deployable in our cloud, your cloud or fully on-prem
- Streams over standard real-time video to any screen
- The same sub-second start-up in every environment
Need it inside your own network? See self-hosting.
Keep exploring
- ProductThrifty Studios Design, voice and deploy an artificial human built on Maya-2.
- WhitepaperArtificial Humans 2026 The enterprise guide to AI humans: what they are, how they work and where to start.
- ResearchMaya-Lite Our compact real-time talking-head model that runs on a single consumer GPU.
- ResearchAll research Our approach, our team and the rest of our published work.