01 · Overview
Lifelike talking heads, without the data centre
The most convincing digital humans usually need racks of server GPUs to run live. Small models run anywhere, but they look like puppets. Maya-Lite, designed in-house at Thrifty AI, is built to have both: a compact model that streams a talking head faster than real time on one consumer graphics card.
That makes it the engine for AI humans at scale: thousands of concurrent support conversations, kiosks in a lobby, or a fully offline deployment on a site with no internet at all.
02 · Positioning
Why compact
Talking-head models usually force a choice. Rig-based models are fast but move a simplified mask of the face. Heavier models paint every pixel, but run far below real time.
We wanted both: a real, pixel-level face and speed to spare, at a size that fits on one GPU.
Every pixel of the face is generated, rather than a simplified rig being moved over a still image, and it all happens fast enough for a live call.
03 · Variants
Swift and Studio
Both variants belong to the same family and share the same look. They differ in the balance between speed and detail, so you can choose per deployment.
Speed first
Maya-Lite Swift
Fastest
- One consumer GPU
Tuned so every frame is cheap to produce. Built for kiosks, high-concurrency support and anywhere latency matters most.
Detail first
Maya-Lite Studio
Sharpest
- Real time on higher-end GPUs
Keeps skin texture, teeth and fine detail. Built for brand presenters and premium experiences where the close-up matters.
Same face, both variants, any language
04 · Streaming
Built for live conversation
Offline video models get the whole script up front. A live AI human doesn't: it has to start moving its lips while the voice is still being generated, and keep going for as long as the conversation lasts. Maya-Lite was built for exactly that situation, not adapted to it afterwards.
05 · Architecture
Architecture
Maya-Lite is a 1.2B-parameter diffusion transformer that generates talking-head video one chunk at a time, straight from a live audio stream. Audio arrives in 1.32-second chunks and is cached with up to 8 seconds of context, so lip-sync stays smooth across chunk boundaries.
The reference portrait is concatenated directly onto the model input, which gives every frame a pixel-aligned anchor against identity drift. Two variants share this design and differ in how finely they compress the video latent: that single choice trades speed against detail.
In
Transformer block × N · 1.2B
- Self-attention, bidirectional inside the chunk
- Reference cross-attention
- Audio cross-attention, multi-layer speech features
- Feed-forward
Out
Video chunkLatent compression set by variant
Lip-sync that keeps up
In a live call the voice arrives a moment at a time. Maya-Lite follows it as it comes, so mouths stay precise instead of lagging or mumbling.
A clean first second
A stream starts from nothing but a photo. Maya-Lite is tuned for that moment, so the opening words look as natural as the rest.
A face that stays itself
Identity holds from the first minute to the last, so the person on screen never slowly turns into someone else.
Two variants, one design
| Variant | Latent compression | Throughput | Hardware |
|---|---|---|---|
| Swift | 32×32×8, one token per 8,192 pixels | 96 FPS | 1× RTX 4090 |
| Studio | 4×8×8, about 32× more tokens | 10.81 FPS | Real time on 2× RTX 5090 |
Swift compresses the video aggressively so every step is cheap. Studio keeps skin texture, teeth and fine detail.
06 · Evaluation
Evaluation
We evaluated Maya-Lite against six leading talking-head models on studio-quality and in-the-wild footage, in both streaming and offline modes.
Quality on studio footage, streaming
| Metric | Swift | Studio | Best streaming peer | Best offline peer |
|---|---|---|---|---|
| Frame realism (FID) | 11.37 | 9.97 | 9.00 | 13.53 |
| Video coherence (FVD) | 126.52 | 111.38 | 155.71 | 113.31 |
| Lip-sync confidence | 4.21 | 5.73 | 4.60 | 5.17 |
| Lip-sync distance | 9.49 | 8.77 | 9.21 | 8.69 |
| Throughput (FPS) | 96.00 | 10.81 | 45.04 | – |
Lower is better for FID, FVD and lip-sync distance; higher is better for the rest. Best result per metric across the models in each group, so no single peer achieves every figure. On in-the-wild footage Studio posts the best lip-sync of any model tested (confidence 5.53, distance 8.01).
Evaluation sets: studio-quality talking faces and in-the-wild faces with varied lighting and pose, in streaming and offline modes. Throughput is measured per GPU on the streaming split. Peers are anonymised groups of six leading talking-head models. Training data: 330,000 clips, 782 hours, 60,000 speakers, 15 languages, 512×512 face-centred crops.
07 · Fit
Where it fits
Maya-Lite is for the places a data centre can't reach, and the places where one GPU has to serve a lot of people.
Kiosks
Lobbies and retail
A face at the front desk that runs on the machine under it, with no data centre behind it.
Scale
High-concurrency support
Many simultaneous conversations from a single GPU, so cost per conversation stays low.
Offline
Air-gapped sites
Everything runs on your hardware, which is what regulated industries ask for.
Edge
Close to the user
Small enough to sit at the edge of your network and respond without a long round trip.
Keep exploring
- ResearchMaya-2 Our full-body flagship model: 0.90 s start-up and 32 FPS.
- ProductThrifty Studios Design, voice and deploy an artificial human in minutes.
- WhitepaperArtificial Humans 2026 The enterprise guide to AI humans: what they are, how they work and where to start.
- ResearchAll research Our approach, our team and the rest of our published work.