Images, motion, and sound: Pruna expands control over video generation with MiniMax H3

Pruna introduces P-Video-2-Pro-Reference, a MiniMax H3-based model combining images, videos, and sounds to guide 5-to-15-second sequences.

Preserving a character's appearance, reusing the movement of an existing sequence, and using an audio track to direct a new video. P-Video-2-Pro-Reference, the new model offered by Pruna AI, combines these capabilities in a multimodal generation system. Based on MiniMax H3, it can leverage up to nine images, three videos, and three audio files as references within a single request. An evolution designed for productions where the output must respect pre-existing elements, rather than relying solely on a text description. Multiple References to Guide a Single Generation

P-Video-2-Pro-Reference relies on a generation principle conditioned by reference media. The user provides a text prompt accompanied by at least one image or video. Audio files are optional.

Images can be used to specify the appearance of a character, product features, an environment, or an artistic direction. Videos provide additional information about movements, framing, and pacing. As for audio references, they allow directing the mood or the relationship between imagery and sound.

The model accepts up to nine images, three video clips, and three audio files. Video references must last between two and fifteen seconds, within a cumulative limit of fifteen seconds. Audio files are also capped at fifteen seconds in total.

These elements are not necessarily reproduced identically. They serve as guidance that the model interprets to generate a new sequence.

This distinction matters particularly for advertising productions or creations featuring recurring characters: a visual reference can promote output consistency without guaranteeing a perfect reproduction of every detail. Transferring Movement and Controlling Start and End Frames

Using reference videos allows directing how a subject moves, how the camera evolves, or how a scene is organized.

Pruna also allows defining a start image and an end image. These two references serve as flexible keyframes to guide the sequence's progression.

They count toward the limit of nine visual references. Their role remains indicative: they do not constitute mandated images with absolute exactness.

The model also features a seed parameter to reproduce certain generation conditions, as well as aspect ratio control. Eight options are provided, including 16:9, 9:16, 4:3, 3:4, and square.

An adaptive mode also allows the model to determine the proportions based on the submitted references. An Audio Track Preserved or Simply Used as a Reference

Audio processing is another key feature of P-Video-2-Pro-Reference.

By default, submitted audio files serve to guide generation without necessarily being kept as the final soundtrack.

With the `keepaudio` option, the user can request that the model retain a specific audio file and generate a sequence tailored to it. This feature requires providing exactly one sound file.

If the recording is shorter than the video, the model generates the remainder of the sequence after it plays. If the audio exceeds the intended duration, it is trimmed.

When an audio file has a duration close to the requested one, with a maximum difference of one second, it can also dictate the length of the produced video.

This operation opens up possibilities for sequences built around a voice, music, or a sound event. However, it does not guarantee perfect synchronization for every situation. Up to Fifteen Seconds at 768p

P-Video-2-Pro-Reference generates 5- to 15-second videos, in 480p or 768p, at a frame rate of 24 frames per second.

Two operating modes are available. Speed, enabled by default, prioritizes execution speed. Cost allows reducing the generation price.

Unlike P-Video-2-Pro, which also offers a Quality mode, this variant with multimodal references does not feature a third quality tier.

An automatic prompt rewriting system, named prompt upsampler, is provided with three settings: Turbo, Max, and Off. It can enrich the description before processing, without changing the price.

When a start or end image is provided, Turbo processing remains applied, even if automatic prompt enhancement has been disabled. Pricing That Includes the References Used

Pruna announces a generation price starting at $0.015 per second at 480p in Cost mode.

The rate depends on the resolution and the selected mode.

Unlike certain services that only charge for the output duration, P-Video-2-Pro-Reference also applies fees for media used as references.

Each image costs $0.005 in Cost mode or $0.01 in Speed mode. Video references are billed based on their duration, starting at $0.03 per second, and audio references starting at $0.0015 per second.

A ten-second video at 768p, generated in Cost mode with three reference images, thus amounts to $0.265, with no other media added.

The final amount is determined after generation, based on the duration actually produced and the references taken into account. The starting price therefore does not necessarily represent the full cost of a multimodal operation. A MiniMax H3 Variant Accessible via API

P-Video-2-Pro-Reference builds upon MiniMax H3, but its interface and parameters are provided by Pruna AI.

It is therefore not a new foundational model trained independently by Pruna. The company adapts H3's generative capabilities into an offering that includes multiple types of references, control parameters, and two pricing modes.

The model is offered via Pruna's API and announced with several inference partners, including Cloudflare, Replicate, Scenario, Together AI, WaveSpeed, and Runware.

Actual availability may vary by platform. Some partners are still announcing integration to come.

Usage is subject to the terms of the MiniMax H3 Community License, particularly regarding permitted content and commercial rights. Reference Consistency Remains to Be Measured

By combining images, videos, and audio files, Pruna addresses a frequent constraint of generative productions: matching multiple elements within the same sequence.

The approach can apply to ad campaigns, product demonstrations, animated