t3版本t3(ver-09 beta) (ID: 175170)

model for asians

描述:

I haven't decided on a finished version yet, so I'll upload a test version first.

▶▶ Stable Diffusion ver1.6 application settings ◀◀

▶ These are the results of the actual picture test.

After testing more than 10,000 pieces by changing the options below by 1, the best was selected.

We will inform you after confirming the figure at which the picture will not be broken.

▶ Sampling method recommendation ( ) is the CFG Scale application range.

Euler(3~26), Euler a(3~22), DPM2 Karras(1~22),

DPM++ 2M SDE Karras (3~25), DPM++ 3M SDE Karras (1~20)

Heun(6~17), Restart(1~18), DPM++ 2S a Karras(1~17)

▶ Hires. fix - Upscaler recommendation

R-ESRGAN 4x+, ScuNET, ScuNET PSNR

▶ Sampling steps: 20~40

▶ Hires steps: 0~20 (no difference)

▶ Upscale by: 1.5~2 (if your PC is good)

▶ Denoising strength: 0.10~0.6 (no significant difference)

▶ Use ADetailer unconditionally

▷ ADetailer model: No difference, use any one

▷ Only one face can be used for positive prom.

▶ Two ways to set CFG Scale

Method-1

▷ Enable Dynamic Thresholding (x)

▷ CFG Scale: 4~7

(If you feel it's a bit strong, turn it down, if that's your style

Turn on Enable Dynamic Thresholding and apply -1 above CFG Scale)

Method-2

▷ Enable Dynamic Thresholding (o)

▷ Mimic CFG Scale: -1 lower than CFG Scale

▷ Mimic Scale Scheduler: Cosine Up 1~3

▷ CFG Scale Scheduler: Half Cosine Up 1~3

▷ CFG Scale: 18~22

▶ Set the rest as you wish.

There may be cases where the picture appears strange due to changes in numerical values.

So don't worry, instead, run a test and find out the setting range.

Please note.

Set the rest as you wish

(Note: When using a detailer, around 0.1 to 0.5 is good. It doesn't matter if you don't use it.)

训练词语: