UGC 偷参考美学：Astra 拆 JSON → 2.5 出片

Prompt

if your ai pr stills still look fake, you’re probably describing the vibe instead of stealing it

here’s what i do with gpt-image 2.5 when i need real-looking ugc photos for a product

chatgpt image 2.5 is currently the best model but with no reference it still goes plastic

so i don’t write “cinematic natural light soft grain” essays

i grab a real photo that already has the look i want

drop it into gpt-6 astra and say

“analyze this photo and give me a very detailed json prompt describing its visual style. break down the color grading, the lighting and the exact colors so i can apply that look to my own images”

astra reverse-engineers the grade for me
i was never typing that accurately by hand

take the json into gpt-image 2.5
attach the product
prompt

“using this json as reference, generate a person holding my product”

then i generate one without the json and compare
the lighting + color grade is the whole difference

when i like the person, i save that frame
every next shot i reattach it and say keep the same face

that’s how i keep one consistent creator across a whole product set
instead of a new random face every time

want a different mood
i just tweak the json in astra and run it again

quick check before you post
flat default colors
plastic grain

fix those two first

json handles the light
character ref handles consistency
you handle the product in frame

that’s an actual ugc stills workflow
not hoping the model feels realistic today

I have created a playbook for it

Comment " workflow" to get the playbook

Attribution

Author

0x_harness

Published

2026-09-11

Original post

View original ↗

Collected from

xianyu110/awesome-gpt-image2.5 · 2098309152141607051 · MIT