Apps & Software

Pika H3 Flex Released for Local and Hosted Video-Audio Generation

· GlobeWire Daily Editorial Desk · Source: github.com

2 min read

Key takeawaysPika has introduced Pika H3 Flex, a post-trained model designed to generate video with synchronized audio across six sampling steps. The repository provides local open-weight inference alongside hosted API capabilities.
Pika H3 Flex Released for Local and Hosted Video-Audio Generation

Image: GitHub · Photo source · Resized and letterboxed to 16:9; no AI generation

Pika has introduced Pika H3 Flex, a post-trained model designed to generate video with synchronized audio across six sampling steps. The repository provides local open-weight inference alongside hosted API capabilities.

Core Architecture and Adapters

Built on MiniMax-H3 components, Pika H3 Flex uses two task adapters:

  • FL2VA: Handles text conditioning and endpoint frames.
  • Ref2VA: Manages multimodal references.

The pipeline encodes prompts and input media to produce video and audio latents before decoding them into an MP4 output. The software automatically downloads pinned base components and adapters.

Input Limits and Specifications

  • Resolutions: Local inference supports native 480p and 768p with mixed stereo audio. Hosted Pika API endpoints extend this to 1080p, 2K, and 4K.
  • Duration: Requests span 5 to 15 integer seconds, mapped to a 17n+5 frame grid at 24fps (for example, a 5-second request yields 124 frames, or roughly 5.17 seconds).
  • References: Ref2VA accepts up to 12 total references (maximum 9 images, 3 videos, and 3 standalone audio clips). Individual video or standalone audio clips must run between 2 and 15 seconds, with combined durations for each category capped at 15 seconds.

Hardware and Installation

Running local inference requires Python 3.12, an NVIDIA GPU with a compatible driver, and system prerequisites such as git, patch, ffmpeg, and a C compiler. The distribution is named `pika-h3-pro`.

Tested cases indicate that a single NVIDIA H100 80GB GPU handles standard configurations. For requests exceeding available VRAM, users can enable `--cpu-offload` to keep transformer weights on host RAM while trading speed for lower peak VRAM usage. Multi-GPU processing within a single request is not implemented.

Licensing and Verification

The adapters fall under the MiniMax H3 Community License Agreement, while the inference and loader code is licensed under Apache-2.0. Creators should consult the official GitHub repository for complete documentation, hardware verification tables, and setup instructions.

Share this article
Read the original source: Pika H3 Flex: video and audio generation inference - GitHub

AI-assisted reporting; source claims are attributed and have not been independently verified.

About the author

GlobeWire Daily Editorial Desk — AI-assisted news based on attributed published sources.

More to read