How to Make an NSFW AI Video: The Practical Steps
By The Fellowi Team · · 8 min read

The short version: start from a still image, keep the first attempt to 4 seconds at 720p on the turbo model, and write the prompt about motion rather than about the scene. That clip costs 500 coins, roughly two euros, and answers the only question that matters early.
Most first attempts fail for three predictable reasons, and none is the model being prudish. Here is the order to do things in.
Step 1: decide whether you have a reference
This is the fork that determines everything else. Image-to-video takes a still you supply and animates it; text-to-video invents everything from a description.
If a specific face or body matters - and in adult work it usually does - use image-to-video. The still fixes the likeness, which is the single hardest thing to control otherwise. Use text-to-video only when you have no reference and genuinely do not mind who appears. The trade-offs are laid out in the comparison post.
You can generate the still first in the image generator and feed it straight in, which is what most people end up doing.
Step 2: write about motion, not about the scene
This is the mistake that wastes the most attempts. With a start image, the model already knows what is in frame - describing it again competes with the picture rather than adding to it.
Write what changesover the four seconds: what moves, in which direction, how fast, and what the camera does. “She turns her head towards the window, slow, camera holds still” is a usable prompt. A paragraph re-describing her appearance is not, and it will come back looking like a slideshow of the still you already had.
Leaving the camera unmentioned is its own failure. If you do not say the camera holds, you often get a drifting push-in that nobody asked for.
Step 3: pick the cheap settings on purpose
Four seconds, 720p, turbo model. 500 coins. This is not a demo tier - it is the same generator, and the only thing you lose is length and pixel count.
Length multiplies cost linearly and teaches you nothing extra about whether the motion works. Resolution changes fine texture, which is invisible while you are still deciding whether the shot is right at all. Going straight to a 30-second 4K clip costs 33,000 coins, about 66 times the draft, and it is the most common expensive mistake here. The full table is in the costs post.
Step 4: expect the vendor check, and know it is free to hit
There are two separate gates and it helps to know which one you have hit.
Our three hard limits never move: nothing sexual involving minors, age-play, or non-consent. Those are refused outright and no phrasing gets around them.
The vendor model runs its own content check on top of that, and it sometimes declines things we would have allowed. That is the rejection most people actually meet. It costs nothing - the coins come back automatically, in full, every time. In the last week every video failure was this, and every one was refunded.
When it happens, changing the framing of the request is legitimate and often works; the vendor check reacts to phrasing as much as to substance. What does not work, and is not worth your time, is trying to route around our three limits.
Step 5: only then spend on the finished version
Once a 4-second draft shows the motion you wanted, decide separately whether it needs to be longer and whether it needs more resolution. Usually only one of those is true.
Two features worth knowing at this point. Audio is generated with the clip and costs nothing extra. An end frame lets you supply a second still, and the render travels between the two - the most reliable way to get a specific movement rather than hoping the model picks one.
The three reasons first attempts fail
- The prompt described the scene, not the movement. Fixed by rewriting around verbs.
- The start image was a good photograph rather than a good first frame. A tightly cropped portrait has nowhere to move. Leave room in the direction the motion goes.
- Too much was asked of four seconds. One movement per clip. Two actions in four seconds produces neither.
What it costs to learn this
Realistically, three or four 4-second drafts before the motion behaves, so around 1,500 to 2,000 coins - six to eight euros - and then the finished clip on top. Rejections do not count against that, since they are refunded.
One practical fact to hold onto: clips are stored for a limited period and then deleted automatically, so download what you keep. Using them commercially is not restricted - we claim no ownership of what you generate. The studio is at Fellowi Video. If you would rather read somebody making the mistakes than a list of them, the account of producing one clip covers the same ground from the other direction, and the uncensored video overview explains where the limits sit.