LemonSlice's co-founder, Lina Colucci, said it out loud. Existing avatar solutions, in her words, "add negative value - they are creepy and stiff." This is an unusual thing for a founder to say, because the standard move is to insist your category is enormous and only getting better. Colucci's move is more interesting: she agrees the category is broken, and then she explains that LemonSlice is not really in it.
The distinction is technical, and it matters. Most avatar tools take a photo and slide a lip-sync layer over the top - the picture stays put, the mouth moves. LemonSlice's model, Lemon Slice-2, generates every pixel from scratch. It is a large-scale video diffusion transformer, roughly 20 billion parameters, in the same family as OpenAI's Sora and Google's Veo3. The difference is specialization: Sora makes cinematic clips of anything; Lemon Slice-2 makes talking humans, in real time.
Real time is the whole game. A pre-rendered avatar video is a finished object - you make it once, you watch it forever. LemonSlice is chasing something stranger and harder: video that is generated on the fly, at conversational speed, personalized to whoever happens to be watching. Feed it a single image and it produces something like a FaceTime call with a person who does not exist, streaming at around 20 frames per second on a single GPU, complete with facial emotion, hand gestures, and whole-body movement.
And it does not have to be a person. The demos include a photorealistic woman, an anime swordsman with red eyes, and a teddy bear in a bow tie. The model animates a cartoon as readily as a human, which is the sort of detail that sounds like a gimmick until you realize it is a product roadmap - every brand mascot, every game character, every drawn face becomes something you can talk to.
