作者 ZMTL (Zaious.)標題 [閒聊] AI:世界名畫時間 Fri Apr 7 16:46:22 2023
https://i.redd.it/ab4aibrw9gra1.png
https://www.reddit.com/r/StableDiffusion/comments/129hhe7/name_that_painting/
Reddit - Dive into anything
500 votes, 77 comments. 195k members in the StableDiffusion community. Welcome to the unofficial Stable Diffusion subreddit! We encourage you to … ...
Workflow :
Use img2img with 0.4 denoising on the largest image your GPU can handle until
you get a composition you like.
Send it to the Extra tab and upscale it using Langzos x2 (or any size you
wish).
Delete any LORAs (or keep 1 max) in your prompt. Then inpaint at 768x768 or
1024x1024 if possible, at 0.4 denoising (with the two inpaint masked only
options selected)
Polish inpaint with 0.25 around the characters,or easier (if your gpu can
still handle the upscaled pic full size), protect details with masks, (like
the hands and faces) and inpaint not masked the whole picture at 0.2
Bonus:
On very small details like hands, the greater the area you select, the more
context SD will get. (It won’t inpaint a feet if you select a torso with),
but the quality may drop if you select too many things
You can also use Photoshop to add a face or hand when you have nothing to
inpaint. The healing brush/clone brush is good to force things.
Edit,: i forgot, after upscaling, put it in img2img at 0.4 another time, it
will fix the pic a lot, try the maximum resolution your gpu can handle
原圖:
https://i.imgur.com/lleBEyI.png
作者說受限於電腦能力沒辦法再更精細了,但我好期待看到一系列的世界名畫XD
那個清明上河圖有沒有人要來一下...
--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 42.77.180.81 (臺灣)
※ 作者: ZMTL 2023-04-07 16:46:22
※ 文章代碼(AID): #1aBzXZA- (C_Chat)
※ 文章網址: https://www.ptt.cc/bbs/C_Chat/M.1680857187.A.2BE.html
→ bnn: 背景的人塞得比原圖高太多了XD1F 04/07 16:49
推 chewie: 猛然一看還真的蠻強大的4F 04/07 16:52
推 reaturn: 裡面角色我看不出來是誰是正常的嗎6F 04/07 16:52
推 finalL: 不就i2i?有很強嗎,光影還大幅劣化7F 04/07 16:55
--