Brief

LemonSlice has released Lemon Slice 2, a 20 billion parameter diffusion transformer model capable of generating infinite-length video at 20 frames per second on a single GPU. The company's API allows users to create and interact with photorealistic video avatars in real-time, aiming to become the dominant form factor for conversational AI. While acknowledging the challenge of the uncanny valley, LemonSlice claims their avatars are best-in-class and can also generate stylized cartoons and animals. The real-time generation was achieved through a causal model, sliding window attention, GAN-based distillation, and various inference optimizations. AI