very strange face render

#23
by Legume4286 - opened

hello I use orange abyss mix and often my result are terrible it’s look like the sampler work bad :
firefox_SoN707Pc4V.gif
i tested all the sampler and only heun work more or less.

i use --no-half --no-half-vae

I have the problem only with your models i tried to re-dl, with et wihout vea, i erase my install of automatic111 etc… nothing change. did you have any idea why ?
some more exemple :

image.png

prompt : (masterpiece:1,2), best quality, portrait,
Negative prompt: (worst quality, low quality:1.4) , (bad_prompt_version2:1), nsfw
Steps: 16, Sampler: DPM++ SDE, CFG scale: 12, Seed: 915860286, Size: 512x512, Model hash: 0fc198c490, Model: AbyssOrangeMix2_hard, Batch size: 2, Batch pos: 0, Clip skip: 2

I'm having similar problems. none of the sampling methods are working right for me in colab it keeps keeping the same image and occasionally bringing up a odd render or just some background without the character, It was no problem yesterday and everything was fine then. the first render the face was like really distorted but sadly didn't save it. I kept testing it with other sampling methods at least the others didn't have distortion just keep same image.

did you use the last version of automatic1111 ? maybe the last version have broke something.

ok following up on my last message I got one image example here I use the prompt :(best quality, 1girl, black hair, red eyes, ahoge, low twin tails, bikini laying on beach towel, highly detailed, smiling, cute, beach)negative prompt: ( lowres, ((bad anatomy)), ((bad hands)), text, missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, low quality, normal quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts)), (monochrome:1.1), limited palette, hands up,)

and this is the recent test it kept the background the same and girl but changing up the girl's swimsuit and gave her a bow. the whole thing is supposed to change but it doesn't .

image.png

image.png

this example here is the only decent output from my early test today before it went even more weird. It's a bit off especially the background and the extra arm but it matches more to the theme I want to a certain extent. Decent compared to the top results. too bad it stopped there :(
does anyone think it is malfunctioning? I hope it will be back to normal soon as this was the first model I was getting better quality results from as I love to make my ocs on there.

image.png

oof same pic copy here's the decent one

image.png

@Legume4286 @meiyume

The issue here seems to be too high of a CFG and too much emphasis on certain words/tokens, the sampler might be a factor as well, I'd recommend sticking to Karras variants. Perhaps something changed making this a bigger issue, but I personally haven't faced it until I used your prompt and settings.

Here are example images after adjustments with changes highlighted, the second one is without bad_prompt_version2, which without seems to result in different, perhaps better, images in some cases. Hope this helps!

grid-0121.png

masterpiece, best quality, portrait,
Negative prompt: (worst quality, low quality:1.4), (bad_prompt_version2:0.8), (monochrome:1.1), (lowres:1.1), nsfw
Steps: 16, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 915860286, Size: 512x512, Model hash: cadf2c6654, Batch size: 2, Batch pos: 0, Clip skip: 2, ENSD: 31337

grid-0120.png

masterpiece, best quality, portrait,
Negative prompt: (worst quality, low quality:1.4), (monochrome:1.1), (lowres:1.1), nsfw
Steps: 16, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 915860286, Size: 512x512, Model hash: cadf2c6654, Batch size: 2, Batch pos: 0, Clip skip: 2, ENSD: 31337

It's working so far thank you so much

YES !!! after testing your prompt, everything work well! THANKS !
IA-created-012520230951-01.jpg

Sign up or log in to comment