Pika H3 Flex: MiniMax H3 in Six Steps Instead of Fifty, a 15-Second Clip With Sound in Seven Seconds

Features

• ✓ Pika's own wording: post-training cut MiniMax H3 from 50 sampling steps to six, so a 15-second 768p clip with synchronised audio finishes in as little as seven seconds of transformer compute
• ✓ Up to 12 reference files per generation, mixing images, video and audio: a character from one image, a world from another, motion from a clip
• ✓ Pika reports #1 among open-weight models and #7 overall on DesignArena's text-to-video leaderboard at publication time, which is that board's own eval rather than a general benchmark
• ✓ Multi-shot narratives, camera direction (overhead open, crane down, whip pan, orbit) plus hard cuts, dialogue and sound can all be described in one prompt
• ✓ Reference-guided editing lets you swap an outfit, a prop or the whole environment while telling Flex what to preserve

How to Use

1. Decide the intent first: one continuous shot or a multi-shot narrative, because that changes how you combine references
2. Prepare the assets: character images, environment images, a motion reference clip, an audio sample, up to 12 in total, and state in the prompt what each one contributes
3. Be explicit about camera language: position, movement, cut points and dialogue each spelled out rather than left for the model to guess
4. Generate a cheap pass to settle composition, then use reference-guided editing to replace only the parts you dislike instead of re-rolling the whole clip