NSFW Daily · 03
Two people: name both, ban a third, keep limbs on their owners
Once one face holds, the next scene is almost always two people. Failure is no longer only drift. Bodies fuse, sexes swap, a third hand appears, a bystander walks in. Write it before you generate: exactly two clearly adult synthetic people; who is above; who is below; a hidden arm returns to its owner.

Where this lesson sits
Daily 02 locked one still and shot 4 seconds of one adult. This lesson adds the second person only after that face holds. Still one inspectable scene, not a film.
Local two-subject contracts repeat the same line: exactly two adults, and they never exchange faces, sexes, or limbs. The Daily turns that into a prompt you can paste tonight. It does not dump experimental XML into the studio.
Today's idea: two people are two identities, not one blob of motion
“Two people holding hands” is not an acceptance test. The model can knead them into one silhouette, put his face on her body, or invent a watcher in the corner. Name identity A and identity B before you generate: hair, sex, which side of the frame, what the primary contact is.
Headcount is a hard constraint. “A couple” is not enough. Write “exactly two clearly 21+ synthetic adults, no third person.” Mirrors, glass, and TV reflections count as a third person until proven empty.
- Write A and B separately: sex, hair, build; they do not inherit each other
- Freeze position: who is above / center, who is below / left; no side swap
- One primary contact — do not stack conversation, hand-holding, and a position change
- No third person, no bystander, no mirror-figure treated as a character
Through occlusion, every limb still has an owner
Anatomy fails after overlap. A hand goes behind a back and comes out on the wrong shoulder; two bodies become three hands or one shared leg. Write that each person keeps one head, one torso, two connected arms, two connected legs; a hidden limb stays with its owner and reappears on that person.
Do not ask them to “share a body,” “switch roles,” or “duplicate limbs.” Those look like style words. They permit fusion.
Easy to fuse
Make a passionate hotel romance scene with two people, cinematic, lots of tangled limbs, extra people around is fine.
Inspectable
Only A and B. Seated side by side, hand-holding, 6 seconds. Hands are countable, faces do not swap, no third person. Hidden arms return to their owner.
Two clearly fictional, unrelated adults aged 25+ sit together in warm window light. One finally offers an apology; the other listens and gently holds their hand. Keep both fully clothed. Frame faces and hands with one continuous camera move, natural dialogue and room tone. Build romance through eye contact, distance and pauses.
How to watch the file
Do not grade a two-shot only on whether the romance feels convincing. Count people, heads, and hands first. Our two-shot sample inspects blocking. It does not prove two-person identity is solved.
- Two people from first frame to last — did a third walk in?
- Did A's face land on B, or did the sexes swap?
- Hand count: a third hand, a detached hand, an arm on the wrong shoulder?
- When a hidden leg or hand returns, is it still on its owner?
- Did room, clothes, or key light swap mid-shot?
- On failure, go back to 4 seconds of one person, or one contact and one camera — do not add a bystander for “realism”
Two stills, locked one at a time
On character reference: one still locks A, one still locks B. Write that <IMAGE_0> is the woman and <IMAGE_1> is the man. Do not collage both faces onto one board. If one still cannot hold A, two stills will borrow each other's faces.
Text-to-video has no identity lock. Use it to try contact and framing. For the same faces, return to consistent character with one holding still per person.
Put the boundary in front of both people
Both characters are synthetic adults, 21+. No minors, including adult bodies that look underage. No unauthorized real-person sexualization. Do not pad an adult act with a passerby, a camera operator, or a person in a mirror.
This is production constraint, not a guide to evading review.
Next lesson
Once two-person contact is inspectable, ask who the clip is for. Daily 04 covers locale as a generation mode: English captions on a xianxia hall are still a Chinese world.
Until then, count the hands on tonight's two people.
FAQ
FAQ
Why not just write “a couple in a hotel”?
That is not identity. The model has no stable object called “couple.” Name A and B, the headcount, and who sits where, or you have nothing to inspect.
Where does the third person come from?
Mirrors, TVs, windows, edge “atmosphere people,” and extra heads or hands after a merge. Write exactly two, and inspect edges and reflections.
Do two-shots require character reference?
If you need the same faces, yes. Text-to-video can probe contact and framing. It will not lock identity. One holding still per person, then the two-shot.
I keep getting extra hands in overlaps. What then?
Stay under 6 seconds, keep one contact, and write limb ownership. If it still fails, shoot a two-shot with no overlap before you add occlusion.
Can I use a photo of a real couple?
Only if both are clearly adult and you hold the rights. Default is synthetic. No minors. Do not leave an unauthorized bystander or operator in an adult frame.
Tonight
Shoot only these two tonight
Open text-to-video, paste the prompt, ask for 6 seconds, and scrub the file counting heads, hands, and whether the faces swapped.