Why pulling the camera back put four faces on his shirt

The first thing everyone gets wrong is framing, and they get it wrong in a predictable direction: too tight. Image models default to a face-filling crop. If your prompt doesn't state a distance, you get a headshot — cropped forehead, no headroom, no room behind the person. It reads as a video call, not a TikTok.

So you pull back. And that's where it gets expensive, because widening the frame doesn't just show more room — it hands the model more canvas to invent on.

What actually happened

We loosened a talking-head prompt from a tight crop to waist-up with the room visible. Two things appeared that nobody asked for. The actor's plain band tee came back with four printed faces across the chest — at a glance, in a three-second scroll, that reads as four extra people. And a single held hairbrush became two mirrored brushes, one in each hand.

Neither is a prompt-adherence failure. The model did what we asked — it filled the frame. We'd given it more frame and no rules about what was allowed in it.

The bans that fixed it

• Only one person in frame — and say that this includes posters, prints, photographs and reflections. The model counts a printed face as a face.
• Plain clothing: no graphics, no text, no faces on any garment. This one clause kills the most common version.
• At most one held object. Hands are where duplication happens; naming the count stops the mirroring.

How to check it in ten seconds

Pull the first and last frame of every render and put them side by side. Then read the frame for content you didn't ask for: garment prints, reflections, duplicated hands, anything with a face on it. Both checks take less time than re-rolling a bad clip.