P-Video 2 Pro: MiniMax H3 Video in Seconds, Not Minutes

Video iteration has a rhythm problem. You write a prompt, wait a few minutes, watch the clip, notice the camera went left instead of right, fix one word, and wait again. By the fifth round you've lost the thread of what you were trying to make. The quality was never the bottleneck. The waiting was.

P-Video 2 Pro from Pruna is a direct attack on the wait. It's built on MiniMax H3, one of the strongest video models of the year, and tuned by Pruna to generate in seconds rather than minutes. Text or a first frame in, video with sound out. It's live on each::labs now. Here's what it is, what the speed actually buys you, and how to use it well.

What P-Video 2 Pro Actually Is

Pruna describes P-Video-2-Pro as its advanced video generation model, "based on MiniMax H3", and the quality tier of its P-Video line. According to Pruna's documentation, it creates new footage from a text prompt plus an optional first frame and an optional last frame, and every output includes generated audio.

The spec sheet is tight. Clips run from 5 to 15 seconds at 24 fps, in 480p or 768p. Aspect ratios cover 16:9, 9:16, 4:3, 3:4, 3:2, 2:3 and 1:1, and when you pass a frame, the image decides the canvas. On each::labs the prompt takes up to 2,000 characters.

Two things it isn't. It isn't an editor: Pruna is explicit that it makes new footage rather than modifying existing video. And it doesn't take audio as input. The sound is generated, not conditioned on a track you supply.

Why a Pruna AI Model Runs on MiniMax H3

Pruna's whole business is making models faster without making them worse. P-Video 2 Pro is that idea applied to a frontier video model: take MiniMax H3's quality, then rework how it runs. Pruna's launch materials name MiniMax's H3 work as the foundation, and that's the right way to think about it. The look, the motion and the audio come from MiniMax. The speed comes from Pruna.

That matters for how you read the output. If you've used MiniMax H3 Text to Video on each::labs, you already know the visual language: grounded motion, stereo sound in the same pass, a strong sense of camera. Our MiniMax H3 deep dive covers that model's strengths. P-Video 2 Pro trades some of H3's ceiling (H3 goes to native 2K, P-Video 2 Pro stops at 768p) for a dramatically shorter loop.

Seconds, Not Minutes

Here are Pruna's own numbers. Its launch announcement puts a 5-second clip at roughly 2 seconds of generation at 480p, and about 4.3 seconds at 768p, in Speed mode. The docs put 768p at roughly 0.85 seconds of compute for every second of video in Speed mode, and about 1.82 in Quality mode. Those are Pruna's figures, so test them on your own prompts, but even with generous slack the difference from a multi-minute queue is not subtle.

Pruna also points to DesignArena results, where it says P-Video 2 Pro ranked third for image-to-video quality and fourth for text-to-video quality, while generating in a fraction of the time of the models around it. Benchmarks are a starting point. What changes your day is the loop.

And the loop is the real story. When a clip comes back in seconds, you stop treating each generation as a commitment. You try the low angle and the high angle. You test three openings. You find out in a minute that the idea doesn't work, instead of finding out in twenty. Speed doesn't just save time. It changes how many ideas you're willing to try.

Speed or Quality: Two Recipes, One Model

P-Video 2 Pro exposes the trade as a single setting. The mode field picks the recipe: speed is the default and the fastest path, quality is slower and aims higher. Resolution is separate, 480p or 768p. Pruna recommends speed mode for fast previews while you test prompts, and quality mode for production output.

That gives you a clean working pattern. Explore at 480p on speed, where generations are close to instant. Once the shot works, lock the seed and the prompt, then rerun at 768p on quality. Because the seed and the prompt carry over, the final looks like the draft you approved, only better.

The other dial is prompt_upsampler: off, turbo (the default) or max. It expands your prompt before generation. Leave it on for short, casual prompts. Turn it off when your wording is the spec, such as an exact line of dialogue, so the model says what you wrote.

Direct the Shot in One Paragraph

With a 2,000-character prompt, there's no room for padding, which is a feature. Pruna's guidance is to state subject, action and scene, then add camera movement, lighting, style and audio as needed. In that order, it reads like a shot description:

A barista in a denim apron pours steamed milk into a latte, forming a leaf pattern. Close up from the side, slow push in. Morning light through a cafe window, warm tones, shallow depth of field. Sound: milk hissing, cups clinking, quiet chatter in the background.

Audio deserves its own sentence every time. The model generates sound whether you ask or not, so naming the ambience, the effects and whether there's music costs you nothing. Pruna also lists dialogue scenes with subtle lip-sync among its recommended uses, so a short spoken line is fair game:

Two friends sit on a park bench at golden hour, medium two-shot, static camera. The woman on the left turns and says "We should have done this years ago." Birds, distant traffic, a light breeze. 16:9.

For camera vocabulary that transfers well across models, our guide to AI video camera movement prompts is a good companion.

First and Last Frames for Shots You Can Plan

When you pass an image, P-Video 2 Pro animates from it. Add a last frame too and it solves the motion between the two. This is the mode for product loops, reveals and any shot where the composition is already approved.

A practical pair: generate a packshot as the first frame, then an edited version of the same packshot with the product rotated and the lid open as the last frame. The prompt only needs to carry what the frames can't:

The jar slowly rotates a quarter turn as the lid lifts and floats upward, a soft light sweep across the glass. Gentle whoosh, a single soft chime at the end.

Remember that the frame sets the canvas, so pick the image size you actually want to ship. If you build this as a chain (still, edit, animate), set it up once as an each::labs flow and swap only the product image for each run.

P-Video 2 Pro vs P-Video 2 on each::labs

Both Pruna models live in the P-Video family on each::labs, and they aren't the same tool. P-Video 2 outputs 720p or 1080p, runs from 1 to 20 seconds (or lets the model choose), offers 24 or 48 fps, has a draft mode for quick previews, and accepts an audio file to condition generation. P-Video 2 Pro is built on MiniMax H3, stays at 480p or 768p, runs 5 to 15 seconds at 24 fps, and generates its own audio.

So pick by job. Need to drive a clip from an existing voiceover or song, or need 1080p out of the box: P-Video 2. Want MiniMax H3's look and motion with the shortest possible wait, especially for cinematic beats, ads and dialogue: P-Video 2 Pro. Many teams will prototype on Pro and finish wherever the delivery spec points, including the full MiniMax H3 Image to Video when 2K matters.

Where P-Video 2 Pro Earns Its Place

Pruna's own list of uses is long: marketing ads, music visuals, narrative content, product loops, training materials, trailers and documentary-style coverage. The common thread is volume with taste. Ad teams can test twenty hooks before lunch. Product teams can turn a catalog of stills into short loops without a render queue. Apps with a "make it move" button can respond while the user is still looking at the screen, which is the difference between a feature people use and one they abandon.

Where it isn't the pick: final deliverables that need 1080p or 2K, clips longer than 15 seconds, or edits to footage you already have. For those, browse the text-to-video models on each::labs and route the job to the right one through the same API.

Frequently Asked Questions

Is P-Video 2 Pro the same as MiniMax H3?

It's built on it. Pruna says P-Video-2-Pro is based on MiniMax H3 and optimized for much faster generation. Expect H3's look and audio at up to 768p, with clips from 5 to 15 seconds, rather than H3's native 2K output.

How fast is Pruna AI's P-Video 2 Pro?

Pruna's launch figures put a 5-second clip at about 2 seconds of generation at 480p and about 4.3 seconds at 768p in Speed mode. Quality mode is slower. Treat these as the vendor's numbers and time a few of your own prompts before you design a product around them.

Does P-Video 2 Pro generate audio?

Every output includes generated audio, and you steer it in the prompt by naming the ambience, effects, music or a line of dialogue. You can't upload your own track to this model. If you need to drive video from existing audio, P-Video 2 accepts an audio input.

Can I use P-Video 2 Pro for image to video?

Yes. Pass a first frame to animate from it, and optionally a last frame to set where the shot ends. The image determines the aspect ratio, and the prompt should describe motion and sound rather than repeat what's already in the frame.