NSFW video API

Each family in the table is a single id that reads both ways: send a prompt alone and it generates the whole clip from your wording, or add a starting frame in image_urls and the same id animates that image instead. There is no separate image-to-video id to remember or to send a picture to by mistake. The family pages show the exact body each model accepts next to its limits: the Seedance API, the Wan 3.0 API and the MiniMax H3 API are three of them, and every other row above links to its own.

A take costs the per-second rate of the chosen resolution multiplied by its duration, and the whole amount is taken when the job opens. If the provider fails the job, those credits come back automatically. Nothing renews unless you opt in: pay-as-you-go credits carry no subscription and no monthly floor, just a balance topped up by card or another method available at checkout; an optional monthly plan is also listed on the pricing page alongside the top-up packages. Test a prompt at the shortest duration and the lowest resolution a model offers, and keep the expensive settings for the take you mean to deliver.

Most comparisons of NSFW video hosting stop at the model list, and the model list is the easy part. Novita, for one, prices its video models by the second or by the clip, yet its Acceptable Use Policy (effective 5 August 2026) forbids sexually explicit output and sets aside no exception for adult use, so an explicit clip falls outside its rules whichever model renders it. Replicate's terms (updated 1 April 2026) draw a different line: they exclude non-consensual nudity and unlawful sexual material, and section 8.6 concedes the services may produce graphic adult output, leaving the customer answerable for it.

The safety-checker defaults we read for fal and ModelsLab on 12 September 2026 sit on image endpoints; the NSFW image API page has that detail. ModelsLab's pricing page does list video models beside its image ones, sold on monthly plans that differ by call volume and concurrency. We did not read either provider's video parameters, so check the schema of the exact video model you intend to call before you build on it.

Asynchronous delivery is normal for clips: Venice's video guide, read the same day, queues a job and has you retrieve it by id, the same shape as the POST-then-poll loop above. What Vellria does not add is an outbound webhook. Nothing calls your server when a clip finishes, so your code keeps asking until the record settles, and there is no SDK; the video endpoints are not OpenAI-compatible either, since the one OpenAI-compatible surface is chat, for the Uncensored Chat model. Where it differs is the meter: credits are prepaid and priced per second of footage at the resolution you pick, taken as the job opens and handed back automatically if the provider fails it, with no required monthly plan and no call quota, though an optional monthly plan is available.

Fictional adult content is allowed. An automatic filter reads the prompt for a minor in a sexual context and refuses what it recognizes with a 400 and the code content_blocked, before any credit moves. Bestiality, incest, non-consensual sex, extreme gore, hidden-camera content and deepfakes are ruled out by the terms, not by that filter. A starting frame you attach is not screened for content: only a file that cannot be read is refused. What's allowed and what isn't has the full policy. A real person used without consent, famous or not, is prohibited by the terms rather than something the endpoint can see, so it is yours to enforce inside your product.

The provider hosting each model also moderates on its own terms. When it stops a render, the job ends as failed with an error code, and the credits return as they do for any other failure. For stills rather than clips, the NSFW image API covers the image endpoints; to make clips in the browser instead of from code, see NSFW video generation.