Hey HN, we're the co-founders of LemonSlice (try our HN playground here:
https://lemonslice.com/hn). We train interactive avatar video models. Our API lets you upload a photo and immediately jump into a FaceTime-style call with that character. Here's a demo:
https://www.loom.com/share/941577113141418e80d2834c83a5a0a9Chatbots are everywhere and voice AI has taken off, but we believe video avatars will be the most common form factor for conversational AI. Most people would rather watch something than read it. The problem is that generating video in real-time is hard, and overcoming the uncanny valley is even harder.
We haven’t broken the uncanny valley yet. Nobody has. But we’re getting close and our photorealistic avatars are currently best-in-class (judge for yourself: https://lemonslice.com/try/taylor). Plus, we're the only avatar model that can do animals and heavily stylized cartoons. Try it: https://lemonslice.com/try/alien. Warning! Talking to this little guy may improve your mood.
Today we're releasing our new model* - Lemon Slice 2, a 20B-parameter diffusion transformer that generates infinite-length video at 20fps on a single GPU - and opening up our API.
How did we get a video diffusion model to run in real-time? There was no single trick, just a lot of them stacked together. The first big change was making our model causal. Standard video diffusion models are bidirectional (they look at frames both before and after the current one), which means you can't stream.
From there it was about fitting everything on one GPU. We switched from full to sliding window attention, which killed our memory bottleneck. We distilled from 40 denoising steps down to just a few - quality degraded less than we feared, especially after using GAN-based distillation (though tuning that adversarial loss to avoid mode collapse was its own adventure).
And the rest was inference work: modifying RoPE from complex to real (this one was cool!), precision tuning, fusing kernels, a special rolling KV cache, lots of other caching, and more. We kept shaving off milliseconds wherever we could and eventually got to real-time.
We set up a guest playground for HN so you can create and talk to characters without logging in: https://lemonslice.com/hn. For those who want to build with our API (we have a new LiveKit integration that we’re pumped about!), grab a coupon code in the HN playground for your first Pro month free ($100 value). See the docs: https://lemonslice.com/docs. Pricing is usage-based at $0.12-0.20/min for video generation.
Looking forward to your feedback!
EDIT: Tell us what characters you want to see in the comments and we can make them for you to talk to (e.g. Max Headroom)
*We did a Show HN last year for our V1 model: https://news.ycombinator.com/item?id=43785044. It was technically impressive but so bad compared to what we have today.
EDIT: Tell us what characters you want to see in the comments and we can make them for you to talk to (e.g. Max Headroom)
Sure, that kind of thing is great fun. But photorealistic avatars are gonna be abused to hell and back and everyone knows it. I would rather talk to a robot that looks like a robot, ie C-3PO. I would even chat with scary skeleton terminator. I do not want to talk with convincingly-human-appearing terminator. Constantly checking whether any given human appearing on a screen is real or not is a huge energy drain on my primate brain. I already find it tedious with textual data, doing it on realtime video imagery consumers considerably more energy.
Very impressive tech, well done on your engineering achievement and all, but this is a Bad Thing.
reply