[ AI Art ] - stable diffusion 討論 分享 02

Disable左Windows setting "Hardware Accelerated GPU Scheduling"未?

holy smoke! challenge accepted!
https://www.youtube.com/watch?v=mX9rVCy6jwc

用paperspace練唔到超過6個鍾,有冇人試過kohya個resume traning係咪得,中間d lr rate點計

自己merge咗個vae檔出黎，用咗佢之後啲顏色會比較高動態啲，冇840000，orangemix個啲咁平淡
https://civitai.com/models/70248/color101-vae






大家可以畀下意見

https://civitai.com/models/47085?modelVersionId=55199&fbclid=IwAR0zVhp5safRPUcdrb4Bk9ifkwWHT_uT6swrZBV1yAoUm8g3uJT6KNtgVU4

*this is LoCon

https://blogs.nvidia.com/blog/2023/05/23/microsoft-build-nvidia-ai-windows-rtx/
On May 24, we’ll release our latest optimizations in Release 532.03 drivers that combine with Olive-optimized models to deliver big boosts in AI performance. Using an Olive-optimized version of the Stable Diffusion text-to-image generator with the popular Automatic1111 distribution, performance is improved over 2x with the new driver.

I updated and tested it. it does looks like faster...
my rtx3060 512x512 step20
from 4s to 2s

巴打用BRA?
BRA好難出到啲咁detail background

it is BRA v5, yes.
and background is hongkong LORA.
https://civitai.com/models/54581/hong-kong-old-district

想問如果想整靚腳 lora
係放腳 (without face ) 落去train ，定成傷條女去train ?

更新完driver, cudnn update到8.9.1，用以下setting gen圖，張4090有33至34it/s

Prompt: solo girl, cafe, drinking coffee, blonde hair, blue eyes, smiling
Negative prompt: bad quality, low quality, worse quality
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3432431734, Size: 512x512, ENSD: 31337

想問如果想整靚腳 lora
係放腳 (without face ) 落去train ，定成條女去train ?

更新咗v1版，顏色勝840000，細節除咗眼之外都好過klF8Anime2
https://civitai.com/models/70248?modelVersionId=83798

macbook M1

Total progress: 100%|████████| 30/30 [00:47<00:00,  1.57s/it]

有無得幫一拳超人變返好畫風

如果個model本身有vae baked仲有無得用其他？
我依家用緊BRA但覺得啲色唔夠靚

由零開始可以去邊到追post?
同埋家陣原生都重係1024*1024max?

理論上任何object 都可放落lora train
那怕係一件衫,一對手, 或者腳, (其實civitai 都可以搵到D參考) .....
當然只係放你想train 既object 落個dataset 度啦, dataset 一定要高質素, 要有唔同角度, close up 佔70%, wide 30%, step 就睇你有幾多source image...

剩係得條腿要正則化train,如果唔係真係得條腿

可以，係VAE到揀返你想用嘅就可以bypass baked-in嘅VAE

唔明，要有腳同整個人？

岩岩入手左張4090想試效能
想問有冇一個網站係有得由頭到尾教點樣自己設定？

你諗下你全部比腳,ai就只係識對腳,佢唔識砌返落人度

所以就應該要用囡囡相做正則化等ai知條腿構圖係點應該係咁

正則化點解？有無英文 keyword ？