Pika launches real-time video chat for AI agents
Pika's new PikaStream1.0 model generates talking-avatar video live on a single H100, so AI agents can hold face-to-face video calls instead of producing clips offline.
Pika announced real-time video chat for AI agents. The feature runs on PikaStream1.0, a model the company calls a "real-time visual engine." According to Pika, it generates personalized video at 24 FPS and 480p on a single H100 GPU, with about 1.5 seconds of end-to-end latency from speech to video. Pika reported these figures itself; the post does not mention independent benchmarks.
Users can try it with their Pika AI Self by inviting it to a Google Meet. Developers of other agents can download a Skill from GitHub. Pika describes AI Selves as persistent agents that carry a user's personality. It says the system keeps each agent's memory and context during a call and can still carry out tasks.
The system has three main parts:
- FlashVAE, a Transformer-based VAE trained from scratch. It decodes one frame at a time using a short, fixed context window, so the cost per frame stays the same however long the video runs. Pika reports decoding at 441 FPS at 480p with 1.1 GB of peak memory. It builds on the company's FlashDecoder work.
- A 9B-parameter Diffusion Transformer, first trained as a bidirectional teacher and then distilled into a causal autoregressive student using what Pika calls optimized self-forcing. This lets it stream video chunk by chunk. Each frame attends to its matching audio tokens for lip-sync, and text prompts can be changed per chunk during a live session.
- Multi-reward RLHF, which tunes for identity consistency, lip-sync accuracy and natural motion. A reference image fed in through RoPE positional encodings keeps the face consistent.
The post does not give pricing, availability beyond Google Meet, or comparisons with other real-time avatar systems.