Pruna P Video Edit API Guide: 7 Real Tests, Exact Costs

Every brand I talk to has the same pile of footage sitting on a drive: one good clip, shot once, that now needs to exist in six colourways, four languages and two seasons. Reshooting is the expensive answer. Regenerating the whole scene with a text-to-video model is the lossy answer, because you get a new scene rather than your scene. The Pruna P Video Edit API is aimed squarely at the gap between those two: you hand it a video you already own and one sentence describing the change, and it gives the same footage back with the change applied.

I ran seven calls against it on Segmind, watched every frame that came back, and checked the billing headers on each one. Here is what it actually costs, what it does well, and the three behaviours in the API that are not in the docs.

What Pruna P Video Edit is

P-Video-Edit is Pruna AI's instruction-based video-to-video editing model. Pruna is a European model lab whose whole business is making models cheaper and faster to run through pruning, compression and quantisation, and this model is the editing member of their P-Video family, alongside P-Video (generation), P-Video-Replace (character swap), P-Video-Animate (motion transfer) and P-Video-Avatar (lip sync).

The contract is deliberately small. You supply one source video of up to 15 seconds and one text edit instruction. You may optionally supply up to four reference images to guide the edit. That is the whole surface. There is no mask, no keyframe, no timeline, no per-object selection. On Segmind it runs at https://api.segmind.com/v1/p-video-edit as a synchronous endpoint, which means no polling: you POST, you wait, and the MP4 comes back as binary in the response body.

The design bet is that most commercial video edits are small and describable. Change the shirt colour. Remove the plant. Swap the product for this other product. That covers an enormous amount of e-commerce and performance-marketing work, and it is work that currently costs a reshoot.

What it actually costs

Segmind lists two rates: $0.045 per second of output video at full quality, and $0.025 per second in draft mode. I wanted to know what "per second" is measured against, because that detail decides your unit economics.

My source clip was 5.00 seconds at 720x1280 and 30fps. Every output came back at 5.04 seconds, and every bill matched the output length exactly:

Billing is on output duration, not input duration, and the arithmetic is exact to four decimal places. There is no minimum charge and no per-call fee on top. A 15 second clip, the longest the model accepts, costs $0.68 at full quality and $0.38 in draft.

The number that matters for a catalogue is the one at the bottom of that: forty variants of a 5 second product clip is about $9 at full quality. That is the price of not booking a studio.

Test 1: the basic attribute edit

The canonical case. One source clip of a model in a plain white t-shirt against a studio wall, and one instruction. I followed Pruna's own prompting advice here, which is to describe the desired final state and then explicitly name everything that must not move.

Left: the 720x1280 source. Right: the edited output at $0.2268. Prompt: "Change only the white t-shirt to deep forest green. Preserve its cut, folds, seams and motion. Keep the person, hair, background, camera and lighting unchanged."

This is the model at its best. The green is consistent across all 126 frames, the shirt keeps its seams and the way the fabric catches light, and the face, hair and background are untouched. There is no flicker on the colour and no halo around the edges of the garment, which is the usual tell of a frame-by-frame edit stitched back together.

One thing to plan for: the output is not the same file geometry as the input. My 720x1280 at 30fps came back as 480x848 at 25fps, on every single call. Nothing in the API schema mentions output resolution or frame rate, and there is no parameter to control either. If you are cutting the result back into a 1080p timeline, budget for an upscale step.

Test 2: is draft mode good enough?

Draft mode is 44% cheaper. The interesting question is what you give up. I ran the identical prompt and seed in both modes and compared them frame by frame.

Left: full quality, $0.2268, returned in 197s. Right: draft, $0.1260, returned in 46s. Same prompt, same seed.

Draft is not a lower resolution. Both came back at 480x848 and 25fps; the saving is in sampling steps, not pixels. Measured structural similarity between the two runs is 0.935, which in practice means the edit decision is the same and the fine texture is softer. Draft smooths skin and loses some strand detail in the hair. On a garment colourway test, at the size these clips actually get watched, I could not pick the expensive one out of a lineup.

Draft was also 4.3 times faster, and much more predictable. Across my runs, draft returned at 9.1 to 10.2 seconds of compute per second of video, a tight band. Full quality ranged from 16.3 to 39.1, a 2.4x spread on identical-length clips. If you are building anything interactive, that variance matters more than the mean.

My recommendation: iterate on prompts in draft, and spend the extra 2 cents per second only on the take you are shipping.

Test 3: reference images

The optional images array is where this stops being a colour-grading toy. You pass up to four reference images, and the model uses them to source the thing being added or swapped in. I gave it a product shot of a red leather biker jacket.

Left: source. Right: the jacket transferred from a single flat product photo, $0.2268.

It carried the asymmetric zip, the lapel shape, the collar studs, the zippered pockets and the waist belt across from a flat catalogue photo onto a moving body, and held them for the full clip. It also invented dark trousers below the crop line, which is a reasonable inpaint given the source never showed them, but it is the kind of thing you want to notice before it ships.

This is the capability that makes the model commercially interesting. One creator video plus a folder of product stills becomes a product-variant generator.

Where to point it: three real workflows

Retail and marketplace listings. One turntable clip of a sneaker becomes the full size-and-colourway grid. The reference-image path handles exact SKU swaps, and a plain instruction handles material changes like smooth leather to suede. At $0.2268 per 5 second variant, a 40-SKU refresh costs less than the photographer's parking.

Performance marketing and UGC. When one creator ad wins, the job is to produce every variant of it before the audience tires: different product colour, different on-screen offer, localised text. Text removal works well, and this is the workflow where draft mode earns its place, because you are generating dozens of candidates and killing most of them.

Furniture, interiors and property. Upholstery colour and material variants from a single walkthrough, plus removing the clutter, cables and packaging that make catalogue footage look unfinished.

Calling it from Python

The endpoint is synchronous and returns raw MP4 bytes, so there is no job polling to write. A 5 second clip takes roughly 45 seconds in draft and one to three minutes at full quality, so set your client timeout well above the default.

import requests

resp = requests.post(
    "https://api.segmind.com/v1/p-video-edit",
    headers={"x-api-key": API_KEY, "Content-Type": "application/json"},
    json={
        "video": "https://your-cdn.com/source.mp4",   # required, max 15 seconds
        "prompt": "Change only the sneaker's smooth leather upper to light-gray suede. "
                  "Keep the sneaker's rotation, background, lighting and framing unchanged.",
        "images": [],              # optional, up to 4 reference images
        "draft": True,             # 44% cheaper, ~4x faster
        "prompt_upsampling": True, # default on, rewrites your prompt internally
        "save_audio": True,        # keeps the source audio track
    },
    timeout=600,
)

if resp.status_code == 200:
    open("edited.mp4", "wb").write(resp.content)
else:
    print(resp.status_code, resp.text)   # 400 validation errors are free

Two parameters deserve a note. prompt_upsampling defaults to true and rewrites your instruction before it reaches the model, which is helpful for short prompts and worth switching off if you have already written a careful one. save_audio defaults to true and preserves the source audio track; my test clip was silent, so I could not verify the behaviour on footage that has real audio.

On prompting, Pruna's guidance matches what I saw: make one principal change per run, describe the final state rather than the action, then list what must stay unchanged. The word "only" is doing real work in every prompt above.

Where it breaks

Pruna are upfront that the model is weak on changes that amount to a new scene, on adding objects that have their own motion, and on camera moves. I tested the camera case directly, asking it to zoom out to a full body shot and orbit ninety degrees to a profile.

Left: source. Right: the camera instruction executed, and the scene drifted with it.

It did not refuse. It zoomed out and it orbited, competently. But structural similarity against the source-preserving edit fell to 0.748, and you can see why: the face is subtly not the same person, the background gained a plant, the wall changed, and the lower body is entirely invented. Ask for a camera move and you cross from editing into regeneration, and you lose the one property you came for.

The honest framing is that this is an edit model with a narrow, well-chosen job. Inside that job it is very good and very cheap. Outside it, you want a generation model instead.

How it compares

Pruna publish a benchmark placing P-Video-Edit against other editing models on Human Elo, price per video second and generation time. I verified the two P-Video-Edit rows against my own bills and clocks; the competitor figures are Pruna's own and I have not independently reproduced them.

The shape of that table is the argument. P-Video-Edit is not the highest quality option; Gemini, MiniMax and Seedance all score higher. It is the cheapest credible one, at roughly a fifth of Seedance 2.5 Turbo's price and a third of Gemini Omni Flash Edit's, while sitting within about 60 Elo of them. For a catalogue job where you are producing hundreds of variants and a human reviews the shortlist, that trade is easy. For a hero spot that airs once, buy the quality.

FAQ

What is the Pruna P Video Edit API used for?

Editing an existing video from a text instruction: swapping garments, changing product colours and materials, removing objects or on-screen text, and replacing items using reference images, all while keeping the original motion and framing.

How much does Pruna P Video Edit cost?

$0.045 per second of output video at full quality and $0.025 per second in draft mode. Billing is on output length. A 5 second edit costs $0.2268 and a 15 second edit costs about $0.68.

How long can the source video be?

15 seconds. Longer clips are rejected with a 400 error before any charge is applied, so a too-long submission costs nothing.

Is draft mode worth using?

Yes, for iteration. It is 44% cheaper and about four times faster at the same resolution, and measured similarity to full quality is 0.935. Use it to find the prompt, then run the final take at full quality.

Does Pruna P Video Edit support reference images?

Yes. Up to four images can be passed in the images array to guide reference-based edits such as swapping in a specific product or garment. In testing, a single flat product photo transferred accurately onto a moving subject.

Can it change the camera angle?

It will attempt it, but this is outside what the model is built for. In testing, a camera move caused the subject's identity and background to drift. Use a video generation model for shot changes.

Worth trying if

You have footage you already own and a list of variants you cannot afford to shoot. In seven calls I spent $1.06 and got clean garment colourways, an accurate product swap from a single reference photo, and a clear picture of the edges. The pricing is exact and predictable, the failure mode on invalid input is free, and draft mode makes iteration genuinely cheap.

You can run it now at segmind.com/models/p-video-edit, with no setup beyond an API key.