Pika Releases a Model That Shoots Video With Sound

Pika Labs announced the release of its open-weight video generation model, Pika H3 Flex, built on MiniMax H3, a video generation model. Pika Labs is an AI video company, and MiniMax is a Chinese AI research group. This video model generates video and audio together, and needs only six sampling steps (sampling steps are how many passes the model makes over itself to generate the video). Every audio element — dialogue, music, sound effects, and so on — is generated separately and can be edited independently. Compared to another model called Fal H3 Max, this one is about 1.8 times faster at 768p resolution. Through its API, Pika offers text-to-video generation starting at $0.01 per second.

For AI avatar videos and similar content, there's no post-processing step after generation, since all the audio elements are generated alongside the visuals. That said, running this model at full power takes fairly serious hardware, since it has 33 billion parameters. So most people will probably use the API at $0.01 per second instead of hosting it themselves. Since the weights are open and available on Hugging Face, the AI model hosting site, anyone with enough compute can skip that $0.01-per-second fee.