Pika-H3-Flex: Generate video and audio simultaneously in 6 steps, with up to 12 reference inputs

Pika H3 Flex is a post-trained H3 model for professional creators, generating video and synchronized audio with just 6 sampling steps. It accepts up to 12 reference images to maintain character or background consistency, allowing fine-grained control over camera work and motion via prompts.

Local inference supports mixed stereo audio video at native 480p and 768p resolutions. Using Pika's hosted API, you can upscale resolution to 1080p, 2K, and 4K, and additionally access audio stem separation and prompt enhancement features.

It is suitable for multi-shot storytelling rather than single cuts, motion graphics, product advertising, and reference-based animation production. Unlike conventional methods, it provides a workflow that precisely controls specific angles or movements (crane, whip pan, orbit, etc.) by combining reference images with text prompts.

It can be loaded and used as LoRA weights in a MiniMax H3-based pipeline through the Diffusers library. It follows the minimax-h3-community-license-agreement license and is utilized in video production environments requiring rapid iteration and precise control.