BanWei

RouWei based checkpoint in my "usual" style.

Usage

• Scheduler/Sampler: Euler A + SGM Uniform

• Steps: 30

• CFG: 7 was used in showcase

• Clip Skip: 2

• Positive prompt: Check RouWei for more info regarding style and so on. My merge probably messed up some usage anyway 😅 (i just prompt without specifics in mind nor quality tags)

• Negative prompt: IllusN3g still does the job :D

The model seems to struggle a bit with hands (RouWei has a similar issue), but i tried to limit this a bit.

Long Story™

There is a new player in town: RouWei is like NoobAI, an heavy finetune of IllustriousXL: 7M pictures dataset, 25 days of 4 x H100 GPU (a quick look at GCP place such an instance like the a3-highgpu-4g at 32k$ a month 😱)

Like some others, i caught wind of it with the WAI merge. I didn't want to once again jump on a new trend, but... since i had LoRAs built in my Illustrious journey, i decided to do a one shot test.

Step 0

Testing and merging AlterBan/BancinIXL/HoJ internal LoRAs (the one i used to create the other checkpoints) into a single 32 Dim LoRA with NO CLIP layers. I ended with 0.6 Alter, 0.45 Bancin, 0.15 HoJ.

Step 1

Merging this LoRA in RouWei v0.7 eps. 0.8 was a final good value after some tests.

Step 2

Adjusting/fixing a bit the result. RouWei is biased toward female. That's an usual problem so, i mixed a bit of AlterBan V3 for good measure, still NO CLIP and a slight adjustment of Contrast/Brightness.

This was done using MBW at 0,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2 for the weights, 0,0,0,-1,1,0,0,0 for the adjusting.

Step 3

Merging my Hand Patch LoRA for good measure, at strength 0.9. This does not fix everything, but should help.

And here we are!

Thanks for reading! ❤

NB: Funny thing. So, i don't know a word of Chinese, but if Google Sensei is not wrong, RouWei may be written "肉味" so "Meaty". And BanWei (so, my usual "Bancin" part mixed with RouWei) may be written "班委" so "Class Committee" ^^;