One Word Kept Putting a Second Man in the FrameRAZDEVAI

Hello,

For a while we couldn't work out why some clips came back with an extra person in them. You'd upload a photo, ask for a woman kneeling and a man standing over her, and you'd get the woman, the man, and a second man standing just behind him doing nothing. Same photo, same settings. Sometimes it happened, sometimes it didn't.

It was the word "a".

Write "a man penetrates her" and the model goes and makes a man, even though there's already one in the photo you gave it. Now there are two. Write "the man", or just "he", and the duplicate stops appearing. We tried the obvious fix first and put "two men, duplicate person, extra limbs" into the negative prompt, which did nothing whatsoever. Changing one article sorted it. I still can't tell you why the negative prompt had no effect, only that it didn't.

I work on an AI Porn generator, which mostly means I spend my day looking at things that came out wrong and working out why. What follows is the rest of that list, and the parts we haven't solved.

English beats your own language, and it isn't close

We've run the same prompt in English and in other languages against the same photo. The non-English versions come back vaguer, and they drop parts of what was asked for. It's not subtle once you've seen it side by side.

This annoys me more than it probably annoys the people writing in. Nobody should have to work in a second language to describe what they want in bed. But that's where these models are, and pretending otherwise costs people credits. Keep the English simple while you're at it. Short sentences describing one action beat a paragraph of atmosphere every time.

Six seconds is the actual constraint

Video generations run six seconds. That number does more damage to people's expectations than anything else in the product.

Six seconds is one action. She's riding him, or he's pulling her hair, or the camera pushes in. It isn't a scene with a beginning and an end. When somebody writes a prompt that starts with her undressing, moves to oral, and finishes somewhere else, the model tries to fit all of it into six seconds and you get a blur of half-poses. The clips that come out right are nearly always the ones that asked for one thing.

Pick the single moment you actually want and describe that. The next moment can be a second clip.

The templates are prompts somebody already got working

There are 64 of them live on the site at the moment, and people treat the list like a menu of effects. They're closer to saved settings. Each one is a prompt and a set of parameters that someone sat and tuned until it stopped coming back wrong, which is worth knowing because the freeform box is where credits actually go missing.

So if you want something the list doesn't cover, open the template nearest to it and look at how it's phrased before you write your own. Short, one action, no story around it. That's most of what the tuning was.

Face swaps live or die on the head, not the face

The swap everybody has seen is the bad kind, where a face gets pasted onto a head and the hairline sits half an inch off. It reads as fake instantly, and it's usually the hair that gives it away rather than the face.

What we ended up doing for ai face swap is replacing the whole head, then forcing it to keep the hair, eye colour and nose from the face you supplied, while copying the head rotation and eye direction from the body photo. That last part matters more than it sounds. If the woman in the original is looking down and slightly left, the new head has to look down and slightly left, or the whole thing reads as a sticker.

It also means a straight-on, well-lit source face beats a better-looking one shot at a weird angle. If a three-quarter shot is all you have, expect to run it twice.

What it costs to find this out yourself

Image jobs run 10 credits and video runs 30, so a bad prompt on video costs three times what a bad prompt on a still does. That's why one wrong article was worth chasing down. At 30 credits a go, "a" instead of "the" is an expensive typo.

If you're going to experiment, do it on stills first. Get the look right on an image at 10 a shot, then take the one that worked and animate it. Most of the people who tell us the tool is useless were writing video prompts blind, six seconds at a time, and never found the cheaper way round.