Pruna AI releases P-Video-2-Pro-Replace: swap characters in videos up to 120 seconds long for $0.05/s

It is a familiar pain point for anyone producing ads, music videos, or social media content: you have captured a great take, but the subject needs to be replaced. Perhaps you need another actor, a different look, or different clothing for a region or brand. Reshooting costs time and money, while a conventional deepfake pipeline turns into a collection of several tools. Pruna AI proposes handling the entire job with one model: P-Video-2-Pro-Replace.

What the model does

You submit the source video and photo references of the new character. The output is the same clip but with a different person: the model transfers that person's appearance and clothing. According to the developers, the motion choreography, camera angle and trajectory, framing, background environment, and original audio track remain unchanged.

Here is a simple example: you have a 30-second clip in which a woman walks down a street and turns her head toward the camera. You upload a photo of another model wearing the desired jacket and receive the same scene, the same gait, the same camera work, and the same street noise, but with a different character. The maximum supported video length is 120 seconds, considerably longer than many alternatives designed for short clips.

Under the hood, the model is built on MiniMax H3. Pruna itself is known for model and inference optimization, so the combination makes sense: take a strong foundation and package it as a convenient service.

Speed, price, and access

Generating 10 seconds of video takes about 1.5–2 minutes. Pricing starts at $0.05 per second at 480p. That works out to about 50 cents for a 10-second clip and roughly $6 for a two-minute video, just for the draft resolution. The author of the original post fairly describes this as expensive, especially when several iterations are needed to select the right references.

Access, however, is not a problem. The model can be run through Pruna's API and sandbox, as well as through Runware, WaveSpeed AI, Replicate, Segmind, Wiro AI, Tellers AI, and Eachlabs. For developers, this means it can be integrated through a familiar provider without setting up dedicated infrastructure.

Limitations and what to examine

For now, these are primarily claims from the documentation. There are no independent comparisons, so several aspects should be tested on your own data:

• how consistently the face holds up during sharp turns and when obscured by hands or objects;
• how well complex clothing is transferred, including logos, fine patterns, and transparent fabrics;
• whether artifacts appear along the boundary between the character and the background when the camera moves;
• how lip synchronization behaves when the clip contains speech, given that the original audio remains unchanged.

Resolution is a separate concern. While 480p is suitable for previews and tests, production work will require higher-priced tiers whose costs were not specified in the post.

There is also an ethical dimension. A tool that can replace a person in any video using a pair of photos is equally useful for advertising and disinformation. Product teams should establish policies in advance for consent to use a person's likeness, labeling synthetic content, and storing reference materials.

Practical takeaway

For product teams and founders working with video content, this offers a fast way to test hypotheses: localizing ads for different audience profiles, trying on different looks, and producing creative variants for A/B tests. A sensible plan is to take 2–3 typical clips, process them at 480p, measure the percentage of usable results, and only then calculate the economics. If only three out of ten attempts are usable, the effective cost of each clip will be three times the advertised price.

For developers, another aspect is more interesting: the model is available through numerous aggregators, making it possible to build an abstraction over providers and select one based on price and queue times. The video-model market is rapidly becoming a service layer in which the winner is whoever packages inference best. Pruna demonstrates this clearly: take a strong open foundation, optimize it, and sell it by the second of video.