What people actually build with a video generation API

Updated 2026-10-17

"Generate video from text" is a feature; a use case is a pipeline with an input, a quality bar, a volume and a place where the output lands. The useful question for a developer is not "can the API make video" but "which input mode, how many retries per accepted clip, and what happens before and after the call". Below are five recurring use cases and what each one tends to demand from the API.

1. Product and e-commerce clips

You already have the asset: a product photo on a clean background. The job is to give it motion — a slow orbit, steam rising from a mug, fabric moving — without changing the product itself. That is image-to-video: pass the photo as start_image_url and write a prompt that describes only the motion and camera, not the product.

• Why not text-to-video: the model will invent a product. For a catalogue, "close enough" is a refund.
• Workflow: batch over SKUs, one job per image, short durations, human review queue before publishing. Expect to reject some outputs, so budget for more than one attempt per accepted clip.
• Pitfall: logos and printed text can drift across frames. Keep clips short and reject on a visual check rather than trying to fix it in the prompt.

2. Social ads and short-form variants

Ad teams want many cheap variants of one idea: different hooks, aspect ratios, openings. Here the model choice is mostly a price/quality trade, and the API features that matter are aspect_ratio (for example 9:16 vs 16:9) and duration_secs.

• Workflow: generate with a cheaper model or lower resolution tier to pick winners, then re-render only the winners on a premium model. See the live table below to compare how much the tiers differ today.
• Pitfall: an unsupported resolution/aspect_ratio combination is silently ignored and the model default is used, never a 400. Check the returned file's dimensions rather than assuming.

3. Previsualisation and storyboards

Directors and agencies use generated clips as moving storyboards: camera angle, blocking, mood. Fidelity matters less than iteration speed. Text-to-video is fine for exploring; once a frame is approved as a still, switch to image-to-video from that frame so shot two continues from shot one. For recurring characters or locations, reference-to-video (several images, optionally video and audio references) keeps identity more stable than prose alone — only on models that document it, so check the model page.

4. Game and app asset prototyping

Teams use generated loops as placeholder cinematics, onboarding animations or store-listing trailers before commissioning final art. The requirement is predictable cost and repeatability: pin the model so the look does not shift under you between sprints, and store the prompt, model id and seed-like parameters alongside the output.

• Pin the model, not necessarily the host. A bare model id routes to the cheapest healthy host serving that checkpoint; a model id with a host suffix is a soft preference with fallback; a provider object with only and allow_fallbacks: false is a hard pin.
• Pitfall: output URLs are for downloading, not for hosting. Copy the file into your own storage as soon as the job completes.

5. Localisation and personalised variants

One master creative, many versions: different backgrounds per region, different products in the same scene, restyled looks. Where the master clip exists, the relevant mode is video editing — input_video_url plus a prompt describing the change. It is mutually exclusive with the other input fields, and not every model supports it; the models that do are listed in VideoRouter's docs.

This is the highest-volume pattern, so it is the one where per-second price differences between hosts compound fastest.

Per second, before VideoRouter's 2.5% platform fee. For tiered models each row compares the resolution tier with the widest host-to-host gap. Built 2026-10-17 from the live catalog.

Choosing the mode from the use case

The mode-by-mode tradeoffs are covered in the mode comparison. If you are deciding whether to run any of this yourself, read build vs buy.

Sizing the volume before you commit

Before picking a model for any of these, estimate three numbers: clips per month, average seconds per clip, and how many attempts you need per accepted clip. Multiply them and you have the billed seconds that matter. Product catalogues tend to be high volume and short; previs is low volume and iteration-heavy; localisation multiplies a master clip by the number of regions. Each profile rewards a different mix, so keep the model id in configuration and revisit it as volumes become clear, rather than hard-coding the choice made during the first prototype.

It also helps to separate the interactive cases from the batch ones. A user waiting in an app needs the fastest model that is good enough; an overnight catalogue run should use the cheapest acceptable one and tolerate a slower queue.

What every use case shares

Every pipeline above is asynchronous: create a job, poll GET /videos/{id} until it is completed or failed, then download. Billing happens once at creation from the requested duration, polling is free, and a job that fails upstream is not billed. Plan for a review step and a retry budget rather than assuming one call equals one usable clip. The quickstart has a working request, and the pricing page shows the current spread across models.

Create an API key and run your own SKUs or prompts through two or three models before committing to one.

Frequently asked questions

Keep reading

• What Is a Video Generation API? How It Works and What It Costs
• Text-to-Video vs Image-to-Video vs Reference-to-Video (API)
• Build vs Buy Video Generation: Hosted API or Self-Hosting?
• Best Video Generation APIs: How to Evaluate Them (Checklist)