AI Girlfriends That Send Videos: What Actually Works
Video is the feature most AI companion apps promise and few deliver. Here is how the three implementations differ, what five seconds of video really costs, and how to test an app before you pay.
Published · For adults 18+
Photos were the last frontier for AI companion apps. Video is the current one, and it separates products much more sharply, because video is expensive in a way photos are not. A picture costs a fraction of a cent to generate. Five seconds of video costs somewhere between ten cents and a dollar depending on the model, and that difference shapes every product decision around it.
The result is a lot of marketing that says video and a lot of products that mean something else by it. This guide explains the three things an app can mean when it offers video, what each one feels like, and the tests that tell you which you are looking at before you subscribe.
Three things video can mean
As with photos, the word covers implementations that have very little in common.
Stock clips are pre-rendered in advance and attached to characters. They arrive instantly and look good, because someone curated them, and they have nothing to do with your conversation. You will see the same clip twice, and you cannot ask for anything specific.
Animated stills take an existing photo and add motion to it — a slow zoom, a hair sway, a subtle breath. These are cheap to produce and they read as a moving photograph rather than a scene. They are genuinely nice for a profile card and disappointing as a reply to a request.
Generated video builds new frames from a reference image of the character plus a description of the motion. This is the only version that can perform a specific act, and it is the one that costs real money to run, which is why it is metered everywhere it exists.
| Approach | Wait | Can it do what you asked? |
|---|---|---|
| Stock clip library | Instant | Only if it happens to be in the library |
| Animated still | Seconds | No — it adds motion, not action |
| Generated video | 1–3 minutes | Yes, within the template or prompt it supports |
Why generated video is short, vertical and metered
Almost every AI companion video you will see is about five seconds long and shaped for a phone. That is not a stylistic choice. Current video models are priced per second of output and their coherence degrades as clips get longer: faces drift, hands multiply, and the scene forgets itself. Five seconds at a vertical aspect ratio is the point where quality is reliable and the price is defensible.
That cost is also why nobody sensible offers unlimited video on a flat subscription. If an app advertises unlimited video, it is either not generating video, generating a very cheap approximation, or planning to throttle you quietly once you cost more than you pay. Metering is the honest design: a subscription for chat and photos, with video priced separately in credits.
The useful question is not whether video is metered but whether the meter is legible. You should be able to see what a video costs before you make it, and see what you have left afterwards.
A rough market rate in 2026: a five-second companion video costs the equivalent of 50 cents to two dollars in credits on most platforms. Anything advertised as free and unlimited is one of the other two implementations.
The consistency problem nobody mentions
The hard part of companion video is not motion. It is keeping the character recognisably herself while she moves. The model is rebuilding her from a reference image many times per second, and small errors compound: her face rounds out, her tattoos wander, her hair changes length, her skin tone drifts a shade lighter.
You can see this most clearly with characters who are not ordinary humans. A character with unusual skin, ears or markings is a stress test, because a model that quietly normalises her into a generic person is telling you it is not really using the reference.
Apps that get this right do something specific: they keep one canonical image of the character, generate every video from that same image, and never re-derive her from the last thing they made. Products that chain each generation off the previous one drift within a week.
- Generate two videos of the same character a day apart and compare the faces.
- Pick the least ordinary character available and check whether her distinguishing features survive.
- Ask for a video and a photo of the same moment, then see whether they look like the same person in the same room.
What to test before you pay
All of this is visible in a first session if you know what to ask for. The goal is to find the boundary between the marketing and the implementation quickly, and the fastest route is to request something specific rather than something impressive.
Ask for a particular action, in a particular outfit, in a particular place. A stock library cannot combine those three. An animated still will give you the outfit and the place, and no action. Generated video will give you all three or fail in an obvious way, which is still more useful than a polite substitution.
| Ask for | Stock clip | Animated still | Generated video |
|---|---|---|---|
| A specific act | Unrelated clip | Motion only | The act |
| Her, not a lookalike | Her, always | Her, always | Depends on the reference discipline |
| Two clips, same scene | Repeats | Consistent but static | Consistent if built properly |
How nsfw.fun does it
On nsfw.fun, video is generated, metered and template-driven. You pick a character, pick a motion — there are dozens, from teasing to explicit — and the video is made from that character's canonical reference image. Clips are about five seconds, vertical, and arrive in a couple of minutes.
Pricing is deliberately boring: a photo costs one key, a video costs ten, and Premium includes a hundred keys a month, refilled monthly, at $19.99 every 30 days or $149.88 a year. Extra keys are available in packs if you run out. You can see the cost on the button before you press it.
Everything you generate is private to your account. Publishing to the public feed is a separate, deliberate action, and nothing you make is posted anywhere by default.
The honest limitation: a template is a template. You are choosing from a catalogue of motions rather than describing arbitrary choreography, because template-driven video is far more reliable than free-text video at this stage of the technology.
Frequently asked questions
How long are AI girlfriend videos?
Almost always about five seconds. Current video models stay coherent for short clips and drift on longer ones, and cost scales with duration, so five seconds vertical is the standard across the category.
Why do videos cost more than photos?
Because they are priced per second of generated output and a clip is many frames. On most platforms a five-second video costs roughly ten times what a photo costs, which is why video is metered separately from a chat subscription.
Can an AI girlfriend send explicit videos?
On adult platforms, yes, subject to their limits. Every legitimate platform refuses real people and anything involving minors, and that is not the same thing as a content filter on fictional adult characters.
Will she look the same in every video?
Only if the platform generates each one from a single canonical reference image of that character. Ask for two videos a day apart and compare — drift between them is the clearest sign a product re-derives the character each time.