Pruna

What is Pruna?

Pruna builds image and video generation models where speed and cost are the point rather than a side effect. The company maintains an open source optimisation framework, and its hosted models are the result of that work rather than a separate product line.

That matters because AI video stopped being a capability problem a while ago. You can generate a clip. The question is whether you can generate ten thousand of them this month without the bill or the latency becoming the reason the feature never ships.

Pruna at a glance

Where the money actually goes in video

AI video is billed per second of output, which means the expensive part is not the video you ship. It is the versions you generated first and threw away.

A finished thirty second clip is rarely thirty seconds of generation. It is a prompt you rewrote six times, a reference image you swapped twice, and a motion setting you only understood on the fourth try. Every one of those attempts was billed at full quality, for output nobody will ever watch.

Iteration does not need production quality. You are checking composition, motion and timing, and all three are visible in a lower-fidelity render. Pruna's approach is to make the exploratory pass cheap and keep the full-quality render for the take you keep.

When should you choose Pruna?

Choose Pruna if:

Your video feature has a discard problem. If you keep one generation in five, the cost of the four you throw away is the line on the bill worth attacking, and it is the one most teams never look at.

You want predictable unit economics. Per-second pricing means you can work out what a feature costs before you build it, which is not true of every provider in this category.

You are comparing video providers seriously. Two Pruna models sit behind the same endpoint as the rest of the catalogue, so running the same prompts through several providers is a string change.

Consider another provider if:

You need a broad media catalogue. Pruna serves one image model and three video models on Eden AI, which is deliberate rather than exhaustive.

You need every call synchronous. P-Image returns in the request, but the three video models run as asynchronous jobs, so your architecture has to have somewhere to wait that is not the user's request thread.

Data residency is a hard requirement. Pruna is a European company, headquartered in Munich and Paris, but a head office tells you who you are contracting with, not where the GPUs are. Verify the processing location directly before treating it as a residency answer.

Pruna pros and cons

Pruna models and capabilities on Eden AI

Available Pruna models

The three video models are billed per second of returned video. P-Video-2 is the quality step up from P-Video, and P-Video-2-Pro sits above it. Note that the Pro tier is not a resolution upgrade: it improves conditioning and quality control rather than definition, so do not choose it expecting a larger frame.

P-Image is priced per image rather than per second, at $0.005, and it returns in the request instead of starting a job.

Image generation

P-Image is the odd one out in this catalogue: synchronous, billed per image, and cheap enough that the per-call cost is rarely the thing you optimise. It matters for the opposite reason to video. Image generation looks cheap per call, which is exactly why nobody measures it, and why it becomes one of the largest lines on the bill the moment a feature reaches real users.

How the asynchronous pattern works

The three video models are long-running, so the call starts a job and returns immediately. You then poll for the result. That is the same pattern as every other async model on Eden AI, so if you already handle one of them, Pruna needs no new plumbing.

What can you build with Pruna?

Video features with a cost ceiling

Per-second pricing turns a vague worry into an arithmetic problem. A five second clip at $0.025 per second costs about twelve cents. Multiply by your expected volume and by your discard rate, and you know whether the feature survives contact with real users before you build it.

Iteration loops that do not cost full price

Separating the exploratory pass from the final render is the single largest saving available on a video feature, and it costs nothing to adopt. Explore on the cheaper path, commit to the better model once the composition holds.

Agents that generate video themselves

Most language models cannot generate video. They can write you a prompt for one, and that is usually where it stops. Eden AI publishes its expert models as tools on a hosted MCP server, so a function-calling model can fire the generation itself, poll the job, and return the finished clip. The agent runs its own iteration pass rather than handing the work back to you.

Why use Pruna through Eden AI?

Comparing video providers properly means running the same prompts through several of them and measuring what comes back. That normally costs an account and a contract per provider, which is why almost nobody does it. Through Eden AI, the provider is a string in the request:

curl -X POST https://api.edenai.run/v3/universal-ai/async \
  -H "Authorization: Bearer $EDENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "video/generation_async/pruna/p-video-2",
    "input": {
      "text": "a slow aerial shot over a coastal city at sunrise",
      "duration": 5,
      "dimension": "1280x720"
    }
  }'

One key, one bill, and the same asynchronous pattern for every video model in the catalogue. Changing which model you are testing is an edit, not an integration.

Frequently asked questions about Pruna on Eden AI