Pika releases weights for six-step H3 Flex video and audio generation

Pika post-trained MiniMax H3 and used Projected Distribution Matching Distillation to reduce generation from 50 steps to six. H3 Flex generates synchronized video and stereo audio from text, starting and ending frames, and up to 12 references. Two task adapters are available under the MiniMax H3 Community License, with local inference supporting native 480p and 768p clips of 5 to 15 seconds.

Visit original source