Six video generation APIs with fewer content restrictions, from self-hosted Wan and HunyuanVideo to Mage, Runware, Replicate, fal.ai and ModelsLab, with prices per clip or second, models and the exact safety settings each one uses.
Self-hosting an open video model like Wan 2.2 or HunyuanVideo 1.5 is the only setup with no platform filter, but you need a large GPU and you run everything yourself.
Mage is the best hosted video API for most developers. It supports NSFW content and runs 12 video models through one API, from LTX Video at about 2 cents a clip to Cherry 2 Pro, its Mage-exclusive flagship video model, which goes up to 1080p and 30 seconds.
Runware's content check is a setting you send with each request, and it runs Wan 3.0 from $0.05 a second at 480p.
On Replicate you can switch off the safety checker through the API, and fal.ai only lets you turn it off with account authorization.
WaveSpeedAI and Novita now ban sexually explicit content, and closed models like Veo stay filtered whichever API you call them through.
We ranked each API on what its terms and safety settings actually allow, which video models it runs, image-to-video and reference support, price per clip or second, and how easy it is to build with.
API | Best for | Starting price |
|---|---|---|
Self-hosted Wan or HunyuanVideo | No platform filter | Free (your hardware) |
Mage | Best overall hosted API | About $0.02 per clip (LTX Video) |
Runware | Per-request safety checks | $0.05 per second (Wan 3.0, 480p) |
Replicate | Switching off the safety checker | $0.05 per video (Wan 2.2 Fast, 480p) |
fal.ai | The widest model catalog | $0.04 per second (Wan 2.2, 480p) |
ModelsLab | Flat-rate unlimited plans | $0.075 per clip, plans from $21/mo |
Open video models run on your own GPU, and tools like ComfyUI let you call them from your code.
What works: Wan 2.2 is released under Apache 2.0, and its makers claim no rights over what you generate. Its TI2V-5B model runs on a 24GB card like an RTX 4090, while the larger A14B models need far more memory. HunyuanVideo 1.5, an 8.3-billion-parameter model from Tencent, runs in about 14GB with offloading. LTX-2.5 has open weights too, free to use for companies earning under $10 million a year. There's no safety checker unless you add one and no per-clip fee.
What doesn't: You handle uptime, scaling and queueing yourself, and video needs much more GPU time than images. Renting a GPU on Modal costs about $3.95 an hour for an H100, and popular GPU clouds like Runpod ban pornographic content, so a rented server can still carry rules. HunyuanVideo's license doesn't apply in the EU, UK or South Korea, and the newest Wan models, 2.5 through 3.0, are hosted only.
Best for: Developers with serious GPUs who want full control.
Mage's API runs the same video, image and audio models as its app, which supports NSFW content, through one asynchronous API.
What works: It runs 12 video models with TypeScript and Python SDKs, ComfyUI nodes and an MCP server, paid in Gems at 1,000 for $1. Open models are cheap: LTX Video costs about 2 cents a clip at default settings and Wan 2.2 about 4 cents. Cherry 2 Pro, Mage's exclusive flagship video model, costs about $0.62 for a default 4-second 480p clip and scales to 1080p and 30 seconds. Cherry and several other models take reference images and Characters you create through the API, so the same face carries across clips. Failed generations are refunded.
What doesn't: Your membership's unlimited generation doesn't apply to API calls.
Best for: Developers who want permissive open and Mage-exclusive video models, plus consistent characters, behind one API.
Runware is a low-cost inference API for image and video models.
What works: Its content check is a setting you send with the request, safety.checkContent, with a fast or full mode, rather than a filter that's always on. It runs Wan 3.0 from $0.05 a second at 480p, $0.10 at 720p and $0.20 at 1080p, and Wan 2.5 at about $0.09 a second at 720p. Its terms, updated in September 2026, ban sexual content involving minors and training a model to impersonate a real person without consent.
What doesn't: Runware's terms don't spell out what adult content they allow, so read them against your use case, and it has fewer video extras, like references and characters, than Mage.
Best for: Developers who want Wan models at low per-second prices with checks under their control.
Replicate hosts thousands of open models behind one API, including several Wan video models.
What works: Models that expose a safety checker, such as SDXL, Flux and Wan 2.2 Fast, accept disable_safety_checker through the API, though you can't turn it off on the website. Its terms ban non-consensual nudity. Wan 2.2 text-to-video Fast costs $0.05 a video at 480p and $0.10 at 720p, and you can run your own models with Cog.
What doesn't: Its acceptable use policy bans "obscene" content without defining adult content further, and popular models can queue at busy times.
Best for: Developers who want a big catalog and a simple switch for the safety checker.
fal.ai is a fast inference platform with most open and closed video models available.
What works: It runs Wan 2.2 from $0.04 a second at 480p, Wan 2.5 up to 1080p, and LTX-2 at $0.06 a second at 1080p, alongside closed models like Kling. Its trust and safety policy describes balancing openness with bans on child abuse material and nonconsensual intimate images.
What doesn't: enable_safety_checker is on by default, and turning it off requires account authorization, so unauthorized requests are always checked. Closed models it hosts keep their makers' filters.
Best for: Teams that want one bill for many video models and can get authorization.
ModelsLab sells API access to open image and video models, including models built for adult content on the image side.
What works: Wan 2.2 and LTX 2.3 cost $0.075 a clip, and its unlimited plan for open-weight models costs a flat $149 a month, which suits high-volume apps. Its pages say moderation follows each account's policy.
What doesn't: There's no free tier, with plans from $21 a month, and its terms date from 2023 and don't address adult content, minors or nonconsensual images directly, so you'll set your own rules.
Best for: High-volume apps that want predictable monthly costs.
Most APIs here explicitly ban sexual content involving minors and intimate images of real people made without their consent, and all of it is illegal in the US whichever API or server you use. In the US, the TAKE IT DOWN Act makes publishing nonconsensual intimate images, including AI-generated ones, a federal crime, and since May 19, 2026 the FTC has enforced the Act's requirement that covered platforms remove reported images within 48 hours. If your app lets users publish video, you need a takedown process.
Mage, for most developers. It supports NSFW content and runs 12 video models, including its Mage-exclusive Cherry 2 Pro, through one API with SDKs, Characters for consistent faces and an MCP server. Self-hosting Wan 2.2 is the only option with no platform filter at all.
Replicate accepts disable_safety_checker through its API, Runware's content check is opt-in, and fal.ai lets authorized accounts set enable_safety_checker to false.
Yes. Wan 2.2's TI2V-5B model runs on a 24GB graphics card, and HunyuanVideo 1.5 runs in about 14GB with offloading, both through ComfyUI.
No. Google's Veo API applies safety filters and limits on generating people, and closed models generally keep their makers' filters whichever platform you call them through.
Open models are cheapest: about 2 to 8 cents a clip on Mage, Replicate or ModelsLab at low resolutions, and $0.04 to $0.20 a second for Wan models on fal.ai and Runware, from 480p to 1080p. Flagship models like Cherry 2 Pro cost more, from about $0.62 a clip.
