PROMPT RECORD图像记录
如何制作看起来不像 AI 的 AI UGC 创作者照片 我经过 7 轮测试生成了约 150 个虚拟创作者。到第 5 轮时,在手机屏幕尺寸下我已经分辨不出来了。第 7 轮才是我真正会用的版本 核...
中文说明
如何制作看起来不像 AI 的 AI UGC 创作者照片 我经过 7 轮测试生成了约 150 个虚拟创作者。到第 5 轮时,在手机屏幕尺寸下我已经分辨不出来了。第 7 轮才是我真正会用的版本 核心玩法:暂停的视频画面 -> 房间自身的光线 > 真实的皮肤 > 手机视频"破坏"处理 第一步 - 打开 https://t.co/tTladQiKyi 并开启一个新对话(内置 gpt image) - 不要要求"一张照片"。要"一个有人为自己日常故事录制的竖屏手机视频中的暂停画面" - 设定取景方式:前置摄像头伸臂距离,或把手机支在约一米远的地方。每张片子选其中一种 - 眼睛看着屏幕、略低于镜头位置,而不是直视镜头。这才是人们面对前置摄像头时真正的样子 - 描述房间自带的光线混合:"灰色的窗光加上暖色顶...
原始 Prompt
how to make ai ugc creator photos that don't look ai
i generated ~150 fake creators over 7 rounds of testing. by round 5 i couldn't spot them at phone size. round 7 is the one i'd actually use
the play: paused video frame -> the room's own light > real skin -> phone video ruin pass
first steps
- go to https://t.co/tTladQiKyi and open a new chat (gpt image is built in)
- don't ask for "a photo". ask for "a paused frame from a vertical phone video someone recorded for their story"
- set the framing: front camera at arm's length, or phone propped about a metre away. pick one per shot
- eyes on the screen just below the lens, not into it. that's how people actually look at a front camera
- describe the room's own mixed light: "grey window daylight plus a warm ceiling light". no golden hour, no "bright and airy"
- put the room in: unmade bed, a charger cable, a mug, a laundry basket. rentals, not pinterest
- describe the skin you want, not what to avoid: visible pores, peach fuzz, a few marks, slight shine on the forehead and nose
- add the small stuff: off-white teeth, frizzy flyaways, little to no makeup
- change the expression and camera every shot. in round 1, 7 out of 10 had the same closed-mouth smirk
the ruin pass
- rounds 1 to 5: shrink to 720p and back up, a touch of blur and sharpening, light grain, save as jpg
- round 7 (better): turn the image into a 1 second video at 720p, then grab a frame. real video compression smears the skin and blocks the shadows the way a paused tiktok does. capcut works for this: export 1 s at 720p, screenshot
the catch
- a clean image is still too clean. the ruin pass is what hid the last ai smoothness
- don't overcorrect. in round 2 i asked for unflattering: yawns, harsh light, food in mouth. it looked real but no brand would use it. aim for "attractive person on an ordinary day"
- fair and east asian skin comes out plastic by default. ask for texture by name
- in round 6, polished "brand" light made every shot look like an expensive camera. phone-video framing fixed it
- mirror selfies break physics. say "the whole photo is one reflection, back of the phone visible, no second mirror"
- gpt image blocks tight full body activewear. add a layer or frame from the waist up
- this is what worked in my tests, not a rule from openai
same creator in every shot
- make a 6-view sheet first (front, side and back, full body and close-up, plain grey background), then attach it to every new scene
- on a 5-frame mock ad the face, outfit and product label stayed the same in all 5
what not to do
- don't copy a real person's face from a photo. make fictional people only
- don't have the creator claim results ("this cleared my skin"). they demo, they don't testify
- don't post without the ai label. switch it on at upload
make it look like a frame someone paused, not a photo someone posed
reply with your niche and i'll send you the exact prompt i'd use for it
bookmark this and open it before your next batch