How to generate product videos with an API, at scale

TL;DR

• A good product ad is not one generation. It is a script, a key frame per shot, several shots that hold the same product, targeted edits, an upscale, and a different export for every platform.
• Building it yourself means more than chaining five to twenty models together. It means selecting the right model at every step, engineering the prompt for each one, and designing the pipeline that holds them together.
• Most of what you spend goes on trial and error before the pipeline works at all: the wrong model picked for a step, the prompt rewritten a dozen times, the shots that do not cut together.
• Beyond calling and prompting the model APIs individually, Runway Dev gives you three ways to skip that work:
  • Model Router: set your rules for cost, quality and latency, and every call is matched to the model that wins that job.
  • Workflows: build your pipeline visually, publish it as one endpoint, call it at scale.
  • Recipes: pipelines Runway has already built and tuned, where one call returns a best-in-class product video.

1. What does a finished product ad actually take?

More than a single prompt. A generated clip is raw material. An ad is a piece of direction: a concept, a sequence of shots that cut together, the product rendered accurately enough to sit next to the product page, lighting and palette that match the brand, and a version cut for each platform it runs on.

Each of those is a decision someone makes, and most of them are made before anything is generated. That is why a single call cannot produce one, and why the teams getting the best results are running a pipeline, not writing better prompts.

2. Why start from images of the product rather than a prompt?

Because a prompt returns a plausible product, not your product. Describe “a matte black cordless kettle with a copper base” and the model invents proportions, a logo and a button placement. For anything branded that output is unusable, and no prompt engineering fixes it, because the model was never shown the thing you sell.

So the pipeline starts with reference images. What matters in them:

• Several angles. One hero shot is not enough. Multiple angles are what let the product hold up when the camera moves, rather than drifting as soon as it turns.
• High resolution, product isolated, unobstructed. Whatever the reference does not show, the model invents. A label too small to read comes back as gibberish text, an edge hidden behind a prop comes back as a shape that does not exist, and anything else in frame risks being treated as part of the product.
• Style references kept separate. Lighting, palette and mood belong in their own reference images, so you can change the campaign look without touching the product.

3. How do you plan the shots before generating anything?

Write the script first, then generate the key frames, then generate video from those frames. Text to video hands the model every decision at once. Image to video hands it one decision at a time, and image models are cheaper, faster and far more controllable than video models.

In practice:

1. Write the beats. Three to five shots, each with a purpose: the hook, the product reveal, a close detail, the product in context. This is a script, not a prompt.
2. Generate a key frame per beat. One image per shot, using the product references and the style references. Iterate here, where a mistake costs an image rather than a video.
3. Approve the frames, then animate them. Each approved frame becomes the starting point of its shot, so the video model is deciding motion rather than deciding everything.

The reason this works is control. Once a frame is approved, the composition, the product and the palette are locked, and the only variable left is movement. It is also where most of the trial and error goes, and images are the cheapest place for it to happen.

4. How do you generate shots that look like one campaign?

Hold the product references constant across every shot and vary only the style and scene references. That single rule is the difference between one campaign and five clips of the same object photographed by five different studios.

Then keep the direction specific: camera move, pace, distance, where the light comes from. “Slow dolly in, shallow depth of field, product centered, late afternoon key from camera left” is a shot. “Cinematic product video” is a wish.

5. Which model should you use for which job?

As of today, these are the models behind the best product ads:

Job	Models
Product frames: photoreal scenes, styling, lighting	Nano Banana Pro, Seedream 5.0, GPT Image 2.5
The shot	Seedance 2, Seedance 2.5, Veo 3.1, Kling o3 Pro, MiniMax H3 Max, Wan 3.0, Gemini Omni Flash, Gen-4.5
Changing one thing in a shot you already have	Aleph 2.0
HDR export for broadcast or a color grade	Runway Ruby

All of these models are available on Runway Dev, alongside all the best models from Runway and other frontier labs.

There is rarely a single model to default to.

• The best one changes with the job: a frame carrying packaging text is a different problem from a photoreal scene, a shot with heavy motion is a different problem from a slow dolly, and a draft is a different problem from a master.
• Best is not only about quality either. The strongest model is never the fastest or the cheapest, so every step is a tradeoff between the three.
• New models ship most weeks, and one release can change the right choice for a single step without touching the others. A hard-coded model id is a judgment with an expiration date.

Runway trains frontier models and serves the best from other labs, so we know which one wins which job, and when that changes. That expertise is packaged into the Runway Model Router. You set the priority: on quality, every request goes to the model that produces the best result for that particular job, and on cost or latency it optimizes for those instead. You choose which models it may use and cap what a single generation can cost, then pass a config ID in place of a model name. Every response reports which model ran and what it cost, and new models become eligible automatically. Choosing the right model for your use case walks through creating one.

6. How do you edit an AI-generated video?

Not by regenerating it. In post-production you change one element and leave everything else alone, and regenerating does the opposite: change the prompt and you get a different shot, with different framing, different motion, and a product that may no longer match the one you approved two shots ago. You lose everything that was right in order to fix the one thing that was not.

In-context video editing closes the gap between the video you already have and the one you actually need. Aleph 2.0, on Runway Dev, changes only what you ask for and preserves the rest: background, lighting, motion, framing and timing. Edit one frame and it carries that change through the video to match. Across a multi-shot ad it applies the edit to the shots that need it, so new packaging lands on all of them rather than one at a time.

Regenerate only when the shot is wrong rather than imperfect, or when the composition itself has to change.

7. What format does each platform need?

A master usually ships as several exports: one per aspect ratio, and one per color space where the destination calls for it.

Platform	Aspect ratio	On Runway Dev
TikTok, Reels, Stories	9:16 vertical	1080:1920, 720:1280
Instagram and Facebook feeds	1:1 square, 4:5 portrait	1440:1440, 960:960, 1248:1664, 834:1112
YouTube, web, connected TV	16:9 landscape	1920:1080, 1280:720

Do not crop between them: cutting a landscape master down to vertical throws away the composition you approved and usually the product with it. Aleph 2.0 is the model for reformatting a master into other aspect ratios: it adapts the content to each one, recomposing the shot rather than cutting the edges off.

SDR covers social and product pages. HDR is what broadcast, connected TV, cinema and any color grade expect, and it is a different export rather than a different encode: higher bit depth, a defined color space, and a container the grading suite accepts. Gen-4.5 and Aleph 2.0 output HDR natively. Runway Ruby converts video from any model, ours or anyone else’s, into 10-bit HEVC in HDR10 or HLG, a BT.2020 ProRes mezzanine, or OpenEXR sequences for compositing. If you are not sure whether you need it, what HDR video is and when your product needs it covers the decision.

8. What does it take to generate product videos for a whole catalog?

Generating one product ad is a pipeline. Running that pipeline for two hundred products changes almost nothing about the model calls and almost everything about the code around them.

• The pipeline becomes an asset you maintain. A key frame and a generation per shot, then edits, an upscale and an export per platform. Every step has its own inputs, limits and output format, and a change to any one of them is a change to your code.
• Retries get expensive. Without checkpointing each step, a failure at the export re-runs the generation you already paid for.
• Every hop produces an asset that has to be stored, referenced and tracked against a SKU, and a fresh generation sometimes needs re-uploading before the next step will take it.
• Every model update is your problem. A better video model shipping is only useful if someone rewrites the pipeline to use it.

9. Can you publish the whole pipeline as one endpoint?

Yes. Workflows on Runway Dev turn a multi-step generation pipeline into a single endpoint you call at scale.

You build the pipeline visually in the Runway Creative App, chaining models, modalities and tasks with no code, then publish it as an endpoint and call it from Runway Dev. However many steps it contains, it is one API call.

What that changes:

• The endpoint never changes. Update the pipeline whenever you like, including swapping in a model that shipped this morning, and every change deploys immediately. Your integration stays as it was written.
• Your creative team can edit the pipeline. It is built visually, so the person tuning the prompts does not need to open your codebase to do it.
• Billing is per step, from your Runway Dev credits. Each generation is charged at its own model’s rate. If a step fails, the workflow stops there: you pay for the steps that succeeded and are refunded for the failed step and everything downstream of it.

10. Is there a prebuilt pipeline for product ads?

Yes. The Product Ad Recipe on Runway Dev runs the whole product video pipeline, built and tuned by Runway, behind a single API endpoint.

The reason that matters is not convenience. A production-grade result depends on model selection, prompt engineering and workflow design, and all three are specialist work. Runway builds frontier models, so we know how they behave, and our technical artists build these pipelines for a living. A Recipe is that expertise packaged into an endpoint.

import RunwayML from '@runwayml/sdk';

const client = new RunwayML();

const task = await client.recipes.productAd({
  version: '2026-07',
  productImages: [
    { uri: 'https://example.com/product-front.jpg' },
    { uri: 'https://example.com/product-side.jpg' },
  ],
  productInfo: 'Organic cold-pressed juice, 12oz glass bottle',
  userConcept: 'Bright, refreshing summer campaign with slow dolly moves',
  ratio: '1080:1920',
  duration: 10,
});

The parameters are the brief, not the pipeline: productImages (one to ten, multiple angles), styleImages for lighting and mood, ratio, duration from 4 to 15 seconds, and audio. The two that carry the creative work, productInfo and userConcept, take plain language: describe the product, then describe the campaign. There is no prompt to engineer and no model to choose.

Runway updates the Recipe as better models ship, so the output improves while your code stays the same. Pin the version with 2026-07 if you need behavior to stay identical, or track unsafe-latest to take every improvement as it lands.

Six other Recipes cover adjacent jobs: Product Swap, Product UGC, Multi-Shot Video, Ad Localization, Marketing Stock Image and Product Campaign Image.

11. What does a product video actually cost?

A finished product ad costs a few dollars to generate. Getting to your first one usually costs a great deal more, because the price of a generation is not the price of an approved asset.

Someone learning video generation on this use case will typically spend somewhere between $200 and $500 getting to their first usable ad. The ad itself is a few dollars of that. The rest is trial and error, generations discarded on the way to it.

The Product Ad Recipe costs 216 credits for the first four seconds at 1080p, then 40 credits per second after that. At $0.01 a credit, a finished ten-second ad is $4.56, as little as 1% of what most people spend getting to their first one.

The difference is not a discount, it is the trial and error you do not pay for. Three things drive it:

• The right model per step is already chosen, so you are not paying a diffusion model to fail at rendering a label.
• The prompt engineering is already done, so you are not buying twelve variations to find the one that works.
• Cost is legible before you call. Runway credits are a flat rate per second of output, so you can compute a campaign’s cost in advance instead of reconstructing it from a token formula afterwards.

12. Which approach should you use to create product videos?

Three approaches on Runway Dev, all of them the same API and the same credits.

• Recipes are generally the way to go, unless you have a very specific or original idea, want to use particular models, or want to trade some quality for cost savings.
• Workflows when you already know which models work best for your use case, how to chain and prompt them, and you want full control of the pipeline to execute a specific idea.
• Calling the models individually, with a Model Router, when you want precise control over cost or generation latency, or want to exclude certain models. Runway picks the model for each call, within the rules you set.

FAQ

Can one model generate a finished product video in a single call?

No. A single call returns one shot at that model’s defaults. A finished ad needs a script, key frames, several shots holding the same product, and an export in the right ratio and color space. The Product Ad Recipe collapses that into one call by running the pipeline for you.

Why does my generated product look wrong?

Because it was described rather than shown. Pass real product images as references instead of describing the product in the prompt, and use several angles so the product holds up when the camera moves.

How do I keep the product consistent across shots?

Keep the product references identical in every generation and vary only the style and scene references. Approve a key frame per shot before generating video, so composition and product are locked before motion is added.

Which model is best for packaging text and labels?

GPT Image 2, available on Runway Dev. It is autoregressive rather than diffusion based, which is why it holds up on typography, logos and dense layout where diffusion models garble small text. Text printed on the product itself should come from your reference images rather than being generated.

Can I change one element of a video without regenerating it?

Yes. Aleph 2.0 edits in context on Runway Dev, changing a color, an element or the lighting while motion, framing and timing stay exactly as generated.

What aspect ratio should a product video be for TikTok, Reels and Instagram?

Vertical 9:16 for TikTok, Reels and Stories, which on Runway Dev is 1080:1920 or 720:1280. Square 1:1 and portrait 4:5 for Instagram and Facebook feeds, 16:9 for YouTube and the web. Generate to each ratio rather than cropping a landscape master, which loses the composition you approved and often the product with it.

How do I export HDR?

On Runway Dev, Gen-4.5 and Aleph 2.0 output HDR natively. Runway Ruby converts video from any model into 10-bit HEVC in HDR10 or HLG, a BT.2020 ProRes mezzanine, or OpenEXR sequences for compositing.

What is the difference between a Workflow and a Recipe?

A Workflow is a pipeline you build and publish as your own endpoint. A Recipe is a pipeline Runway built and maintains. Both are one API call.

What happens if one step of my Workflow fails?

The workflow stops at that step. You pay for the steps that succeeded and are refunded for the failed step and everything downstream of it.

How long should a product video be?

Most social product ads run under 15 seconds. On Runway Dev the Product Ad Recipe generates 4 to 15 seconds and defaults to 10, and each video model has its own duration limit per generation.

How do you keep token spend down when generating video?

Video models are billed per second of output rather than per token, so the equivalent of token efficiency here is not paying for generations that never ship. Draft on cheap models, approve a key frame before animating it, edit in context rather than regenerating, and use a Recipe where the pipeline is already solved.

How do you generate a product video with an API?

Send product images and a short brief to the Product Ad Recipe on Runway Dev, and it returns a finished ad from one call. To build the pipeline yourself: generate a key frame per shot from your product references, animate each approved frame with a video model, edit in context where a detail is wrong, then export one file per platform.

How much does it cost to generate a product video?

On Runway Dev the Product Ad Recipe costs 216 credits for the first four seconds at 1080p, then 40 credits per second, and credits are $0.01 each. A finished ten-second ad is $4.56. Building the same pipeline by hand usually costs $200 to $500 in trial and error before the first usable ad.

How many product images do you need?

One to ten. More than one matters: multiple angles are what keep the product accurate when the camera moves. Use high-resolution images with the product isolated and nothing else in frame.

What is a Recipe?

A Runway-built API endpoint that runs an entire pipeline for one use case. You send inputs and it returns a finished asset. Recipes cover product ads, product swaps, UGC video, multi-shot video, ad localization, marketing stock images and campaign images.

Can you generate product videos for a whole catalog?

Yes. Call the same Recipe or Workflow endpoint once per SKU, queue the jobs and store each task id against the product. The model calls do not change with volume, the orchestration around them does.

Will my results change when Runway updates a model?

Only if you want them to. Pin a Recipe to a dated version such as 2026-07 to keep behavior identical, or use unsafe-latest to take each improvement as it ships.

How do I upscale AI video to 4K?

Two routes on Runway Dev: generate at 4K directly with a model that supports it, or generate at 720p or 1080p and upscale afterwards. Upscaling raises resolution and sharpness only, and changes nothing about content, motion or duration. Generating at 4K costs several times the 1080p rate, so upscaling the approved shot is usually cheaper than rendering every draft at 4K.

Can I use AI-generated product videos commercially?

Yes. Runway’s Terms of Use state that Runway does not claim ownership of your inputs or outputs and does not restrict commercial use of your outputs, subject to your compliance with the agreement and the Usage Policy. Worth knowing: outputs may not be unique, and similar outputs can be generated for other users. If you resell access through your own product, each of your end users has to be bound by terms no less protective than Runway’s. Can I use AI image and video generation commercially in my product? covers ownership, input rights, training and disclosure in full.