我以前常用 Midjourney 生成这类网红图，但说实话 GPT 这些天真的非常惊艳。这就是我正在使用的提示词：

I used to generate some of these influencers with Midjourney but GPT is honestly is amazing these days. Here is the prompt I am using: 

PPP = primary subject prior (socially camera-aware, fashion-forward, high pose literacy)
B(x,y)B(x,y)B(x,y) = urban background manifold
G(t)G(t)G(t) = gaze state, where gaze may briefly intersect lens but is not locked as a studio pose
M(t)M(t)M(t) = body motion trajectory during a natural walk / turn / gesture
F(z)F(z)F(z) = depth-selective focus field
L=Lsky+Lbounce+Lstorefront+LpracticalL = L_{sky} + L_{bounce} + L_{storefront} + L_{practical}L=Lsky​+Lbounce​+Lstorefront​+Lpractical​
Image objective:
Capture a frame that lies between candid and performed spontaneity. Subject is the saliency maximum, but the scene must imply real public space, incidental passersby, and unforced timing.
Derived consequences:

body asymmetry > perfect symmetry
one gesture is transitional, not held too cleanly
wardrobe reads intentional, but pose reads discovered rather than staged

background figures, signage, reflections, or traffic should support “lived-in city”
shallow depth of field isolates the subject while preserving enough context to identify the street-fashion environment
Consistency test:

if subject styling is editorial, environment detail must still feel incidental
if gaze meets lens, expression must remain fleeting rather than fully posed
if motion exists in limbs/hair/coat, it must correspond to step timing and not random blur

FORBIDDEN:
full studio pose in a supposedly candid setup

background emptiness that kills the candid illusion

impossible mix of totally frozen coat/hair with obvious walking stride
over-clean symmetry that reads as staged campaign photography