UGC 自然皮膚翻車：2.5 在這場景偏 mid

提示詞

I spent the last 3 days testing GPT Images 2.5 for AI UGC models. 🤯

not gonna sugarcoat it - it's mid. actually kind of bad for this specific use case.

here's what happened:

i needed a UGC-style creator photo. natural skin, real lighting, someone who looks like she just picked up her phone and started filming - not a "generated" face.

GPT Images 2.5 kept giving me:
- plastic, over-smoothed skin
- that same "AI face" symmetry
- lighting that looks studio-lit even when i asked for natural window light
- hands and props (rings, necklaces, furniture) subtly glitching
i regenerated probably 40+ times. same issues every time. so out of frustration i ran the exact same prompt through Nano Banana Pro.
first output. this 👇

skin texture, pores, real shadow falloff from the window, a necklace that actually sits right, hair flyaways - the kind of imperfect detail that makes a photo feel real instead of rendered. The only is the voice, but i keep working on it.

this is the difference between "AI generated" and "AI that could pass as a real creator."

if you're building UGC ad creative and still defaulting to GPT Images, you're leaving quality on the table. switch and see for yourself.

中文翻譯

過去 3 天我一直在用 GPT Images 2.5 測試 AI UGC 模特。🤯

不兜圈子——它偏 mid，對這一個具體用例來說其實有點糟。

事情是這樣的：

我需要一張 UGC 風格的創作者照片。自然的皮膚、真實的光線，看起來像她剛拿起手機就開始拍——而不是一張“生成感”的臉。

GPT Images 2.5 一直給我：
- 塑料感、過度磨皮的皮膚
- 那種千篇一律的“AI 臉”對稱感
- 即便我要求自然窗光，打光看起來仍像影棚燈
- 手和道具（戒指、項鍊、傢俱）出現細微的錯誤
我大概重新生成了 40 多次。每次都是同樣的問題。於是出於沮喪，我把完全相同的 prompt 丟給 Nano Banana Pro。
第一次出圖就是這樣 👇

皮膚質感、毛孔、窗戶帶來的真實陰影衰減、一條佩戴位置終於正確的項鍊、髮絲碎髮——正是這種不完美的細節，讓照片顯得真實，而不是渲染出來的。唯一的問題是聲音，不過我還在繼續打磨。

這就是“AI 生成”和“AI 能冒充真實創作者”之間的差別。

如果你在做 UGC 廣告素材，卻仍然默認用 GPT Images，那你就是在白白丟掉質量。換一個試試，自己看。

溯源

原作者

ItsNazar17

發佈日期

2026-09-11

原始出處

查看原帖 ↗

收錄來源

xianyu110/awesome-gpt-image2.5 · 2098396894082843074 · MIT