Pika H3 Flex keeps the soundtrack editable

When a generated clip’s music overwhelms its dialogue, a single mixed soundtrack leaves little room to fix the balance. Pika H3 Flex, announced October 29, offers five separate audio stems through Pika’s API: dialogue, score, sound effects, foley and ambience. It also accepts a combination of image, video and audio references. The useful change is room to revise the edit after generation. Pika’s announcement.

A product image can define the object while a movement reference suggests the performance. The inference documentation allows 12 reference files in total, with separate caps of nine images, three videos and three standalone audio clips; at least one image or video is required. More references are not proof of better control, so start with only the inputs whose jobs you can explain.

Choose hosted features or local inference

The downloadable adapters produce native 480p or 768p video with mixed stereo audio. Pika’s higher-resolution finishing, separate stems and prompt enhancement are hosted additions. A 4K export therefore should not be mistaken for native 4K generation in the local package.

The model card also separates licensing: model derivatives use the MiniMax H3 Community License Agreement, while example code is Apache 2.0. The code’s licence does not extend to the weights or example media. Review the applicable model terms before building a distribution around them.

Pika’s API pricing lists a $10 monthly membership with $10 in API credits each month. Usage is then charged by the selected operation: 480p text-to-video starts at $0.01 per output second, while reference-to-video starts at $0.03, with video-input charges listed separately. Higher resolution, stems and other operations can add cost. These are API prices, not a verified creation-app subscription entitlement.

Keep a useful draft before finishing it

Pika’s seven-second headline measures transformer computation in a controlled vendor test, not the entire wait for an API result. Its post scheduled creation-product access for November 2; account rollout was not verified here.

For a proposed first exercise, use one short scene and one clearly assigned reference, then check movement and speech before paying for high-resolution finishing. If the hosted stems are available in your workflow, reduce the score and listen for whether dialogue remains understandable. This exercise was not executed for this report.

The AI video tool guide helps compare workflows. For multi-scene planning and generation quotes, see PixelDojo’s film update.