[ AI Art ] - stable diffusion 討論 分享 01

會
我行 4090 desktop 9秒gen1張128 step 圖
我仲用自己女神 checkpoint 、自己女神Lora 同textual inversion embeddings

desktop RTX3060 - 12g

512*512
euler A , 20 steps
3 second.

RTX4090

[ TEXTUAL INVERSION ]
so chill

https://civitai.com/models/167/poolrooms

liminal space

Webui Lora右上角有個 "!" ，
click 入去可以睇到Lora 資料，例如trigger words, Training 用嘅model之類

Thanks!
呢個 extension都超好用, 自動攞全部checkpoint, lora...預覽圖，顯示trigger words等等
https://github.com/butaixianran/Stable-Diffusion-Webui-Civitai-Helper

宮崎駿 Lora 好正

parameters

ultra high res,detailed shadow, best quality, depth of field,1girl,(16-year-old girl),portrait,front view half body,
,(angry:1.2),(street),short hair, <lora:studioGhibliStyle_offset:1.2>,assault rifle:1.2,military uniform, military hat, close up shot,screaming,
Negative prompt: paintings, sketches, (worst quality,low quality,normal quality:1.9), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans,nipples, watermark, signature, text,(bad finger),bad anatomy, bad hands,(six_fingers),(nail_art),nail polish,(multiple views:1.5),(hat)
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 2752815644, Size: 512x768, Model hash: 7f96a1a9ca, Model: anythingV5Anything_anythingV5PrtRE, Denoising strength: 0.5, Hires upscale: 2, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B

用之前漁夫巴打個 Notebook 彈出句咁既野，但我咩到冇做過，剩下放左d新既 Lora 入 Gooogle Drive

打入 GPT 問之後打跟佢打呢兩句 Code 又用到了，究竟點解

Chat GPT 幫到手

呢個次序睇落好啲

Checkpoints,Lora,Textual Inversion,Hypernetworks

邊度係Webui Lora右上角搵唔到個感嘆號

之前Run 漁夫個 look book 冇問題，但昨天 run 既時候會出現
RuntimeEror: Detected that PyTorch and
torchvision were compiled with different ClDA versions. PyTorch has CUDA
Version=11.7 and torchvision has CUDA Version = 11.8 . Please reinstall the torchvision that match PyTorch install. 即使用一個新 Google Account 都解決唔到

extreme skin details

4X UltraSharp (left) 4X R-ESRGAN (right)
upscale using 4X UltraSharp is much better!

有無高手答到以下問題：

Dream booth 如果 train 200張相會唔會太多？
Dream booth 應該set 幾多training step per image ?
我set 100, 120, 150 又好似無大分別，train 出黎個女神都好似麻麻地似，又驚 train 過多steps 會 overtrain 個model

Train 女神個 checkpoint 係咪 chilloutmix ni 最好？用 realistic vision v2 train 出黎好鬼妹樣
想 train 亞洲人

同樣想問Lora
用幾多張相 train ? 200張會唔會過多？
Lora train 係咪應該用大頭相好d ? Training steps 應該幾多次per image ?

lora好似好過db好多
我嘅經驗相嘅數量唔係重點，重點係佢嘅質素
步數我見有啲人分享2000步做總數，行幾轉epoch
亞洲人面孔就chilloutmix做base

以為train 多d準d

我之前都係咁理解，事實上我依家用緊嗰個都仲係200步數/pic嘅lora
但有講過話過度training會令到caption fixate咗某個印象，個lora嘅靈活性會低咗，意思係冇咁容易去換衫、換髮色之類

但200/pic嘅lora明顯係好過其他嗰啲
我未試過multiple epoch batches，下次試下睇下低步數多epoch係咪就係王道

始終上網揾關於練lora嘅資料都好散亂，無咩人歸納個結論，都係靠自己試

Exactly
就連YouTuber 都無一個 conclusion

咁有無試過textual inversion ? 你覺得點？

未試過，但用textual inversion個感覺冇lora咁高靈活性
例如我用其中一個人哋整嘅ti，我覺得佢個diversity冇咁高
我唔知係個training data低定係點，但同一個對象唔同人整嘅ti好似都係咁
所以我諗緊好似真係佢個diversity冇lora咁高