这两天铺天盖地的福利姬风,核心是低质量自拍。
比如:炫光漏光,较暗的环境,镜子上的污渍,廉价俗气的服饰,杂乱模糊的背景等……
一句话,就是用缺陷去中和AI的过度审美,以增加画面真实感。
LLM系较好的原因是很容易通过关键词“平庸”达到要求,MJ等传统扩散系则要用精确提示词去降低审美,所以相对难些。
Midjourney
![]() | ![]() |
|
|
![]() | ![]() |
|
|
FLUX(Krea Optimized)
Krea 内嵌的Flux尺度比MJ宽得多,虽然也有一定的审核 ,但一些擦边的情况还是很容易得到的。
![]() | ![]() |
|
|
![]() | ![]() |
chinese coser , athletic , slightly muscled, taking tiktok selfie in a sea side assort wooden room through the dusted mirror, you can see the beach through the window, low angle, lolicore, breakcore, cute face, full body, white bikini, web cam, iphone 13,low quality photo,active pose, plain aesthetic ,webcam, background movenment blurry, slightly dusted unclear mirror glass, , clean but messy room,low angle view, focus on the legging as product advertise, cartoon phone shell | chinese coser , taking tiktok selfie in own room through the dusted mirror, low angle, lolicore, breakcore, cute face, full body, unbuttoned white shirt, short pink jeans hot pants , web cam, iphone 13,low quality photo, plain aesthetic ,webcam, background movenment blurry, slightly dusted unclear mirror glass, , clean but messy room,low angle view, focus on the legging as product advertise, cartoon phone shell |
豆包
相对来说,豆包作为LLM式的绘图,可以很好的理解提示,并用更直接更细致的方法去描述需求,所以出来的图会比普通的diffusion模型要精确得多。
![]() | ![]() |
|
|
相关链接:
立即体验