NSFW Daily · 02
Same face: lock one still before you add a second person
Daily 01 collapsed the film into one inspectable scene. Daily 02 is why the face leaks. The usual failure is seven stills at once — profile, smile, other clothes, other light. The model cannot tell which file is the identity. Give it one clean, same-wardrobe, unobstructed still, then shoot 4 seconds.

Where this lesson sits
Daily 01 is what to shoot: one 4–15s scene, five fields, inspectable. This lesson is how to leak less face. Chinese, English, and Japanese are the same class, not empty mirrors.
The consistent character manual records the studio contract: at most 7 stills, optional voice, 720p cap. The Daily does not repeat those numbers. It leaves a shooting order you can use tonight.
Today's idea: the still locks identity, not pose
A reference still can constrain face, hair, apparent age, build, and skin. It is not a contract for pose, camera, room, or wardrobe unless you write the clothes again. Treat the still as a storyboard and the model will drag the portrait stance, the backdrop, even another outfit into the bed.
This is an identity contract, not marketing: the picture defines who. Pose, blocking, and contact are written separately. Tattoos, watermarks, captions, and another person's marks are contamination. Do not let them transfer onto tonight's character.
- Lock: face, eyes, hair, age floor, build, headcount
- Write separately: pose, room, key light, action, camera
- Discard: the still's stance, backdrop, watermark, someone else's marks
One still first, not seven
The API will take up to seven references. Production uses one for the first shot. Seven files with mixed light, hair, and makeup tell the model that all of those people count as her. Run 4 seconds from one unobstructed, same-wardrobe still whose light is close to the clip. Add a profile or a full-body later if you must.
Voice is not a face. Audio without a visual identity still gives you someone else talking.
- Full face, no mosaic, no sunglasses covering the landmarks
- Hair and clothes match tonight's clip; do not pair a day shirt with a night dress
- Clearly adult synthetic character; a real person needs rights you hold
- Repeat short hair / gold dress / one person in the prompt, and name <IMAGE_0>
4 seconds is easier to inspect than 15
Identity drift shows up when action, occlusion, and duration stack. Four seconds is enough for one turn, one blink, one hand on a chair. Fifteen seconds often invents a cut, a new expression, or a second person. Pass the short clip before you add length.
Our public sample is one still, 4 seconds, 3:4, 480p. The gold-dress identity holds in a short shot; face and pose still drift slightly. That is a measured fact, not “consistency solved.” Grade your file the same way.
Hard to lock
Use these seven photos to make a 15-second cinematic hotel romance video. She should be beautiful from every angle.
Inspectable
One still, one person, one room, one action, 4 seconds. Scrub first frame to last and ask only: is it still her?
Use <IMAGE_0> as the only identity reference. The same clearly 21+ synthetic adult woman: short hair, gold dress, same build. The still locks face, hair, apparent age, and body — not pose, room, or camera. She stands in a warm hotel suite, turns slowly toward the lens, blinks once, and rests one hand on a chair. Locked camera, one continuous shot, 4 seconds. Breath and fabric only. No speech, no captions, no second person, no other face.
How to watch the file
Do not judge the poster loop. Open the file and scrub it. On failure, remove variables. Do not add a longer story, and do not stuff five more stills.
- First frame and last frame: the same person?
- Hair, gold dress, build — did they swap mid-shot?
- A second person, a caption, a watermark?
- Is the action still turn / blink / hand on chair, or did a sex scene appear on its own?
- On failure, use a cleaner still, stay at 4 seconds, or lock a still before animating
Put the boundary in front of the still
Default: synthetic, 21+, legal source. No minors, including adult characters who look underage. No unauthorized real-person sexualization. A “true face” under a mosaic is generative fill, not recovered truth.
A reference is a condition, not a face ID. This lesson does not teach you to evade review. How you use the output is on you.
Next lesson
Once one face holds, talk about two. Daily 03 covers naming two adults, banning a third person, and keeping limbs on their owners through occlusion.
Until then, shoot the clip above. If one still cannot hold, two stills will swap faces faster.
FAQ
FAQ
Why not use all 7 reference slots?
The cap is not a recommendation. Mixed light, hair, and clothes average into someone else. First shot: one clean, same-wardrobe still.
Does a reference guarantee the same face?
No. It is a condition, not verification. Occlusion, profile, new light, and extra duration all leak. Our 4-second sample records slight drift.
Can I send voice without a picture?
Voice cannot replace visual identity. For the same face, start with a legal still. Add a voice after the face already holds.
Why 4 seconds instead of 15?
A short continuous shot is easier to inspect for a face swap. Fifteen seconds invents cuts, people, and new expressions. Pass short, then lengthen.
Can I use a photo of a real person?
Only if everyone is clearly adult and you hold the rights. Default is synthetic. No minors. No unauthorized real-person sexualization.
Tonight
Lock this face tonight
Open character reference, upload one legal 21+ synthetic still, paste the prompt, ask for 4 seconds, and watch the whole file before a second shot.