The "far" eye in profiles are suddenly messed up.

[MancV21]

Hi. Everything seemed fine the other day but today when generating faces (with or without face restoration using both methods) the far eye is messed up. Either it's just deformed or makes them look cross-eyed. You can see what I mean from the images I've added below. Sometimes it can be subtle, sometimes extreme, happens with all models too.

Anyone else getting this lately? I thought of maybe resetting the settings for all of the WebUI but can't find an option to do that, is there a way?

EDIT - I REMOVED THE CONFIG.JSON FILE TO RESET THE WEBUI, DIDN'T CHANGE ANYTHING.

Any help appreciated.

[lvcju]

Did you manage to find any information on the problem? I'm having the same issue for months, couldn't solve yet.

[freecoderwaifu]

Try a different negative prompt, really short ones with one or two negative embeddings work better than long schizo prompts. Try without masterpiece, best quality, etc. Easiest way to find a better short negative prompt is going to Civitai and looking at the example images there.

Try different samplers and more steps. Easiest way to determine the optimal amount is going to Live Previews in Settings and set Live preview file format to PNG, Live preview display period to 1 step, Live preview method to Full and Progressbar and preview update period to 1. Then in the txt2img regular generations settings crank up the steps slider to max, see when the image stops changing and converges and details start getting diminishing returns, and adjust the steps slider accordingly. That goes for hires fix steps too, the default 0 = the same as the regular steps, you usually only need a fraction of that.

Keep Sampler Parameters to defaults, in case you changed them after following some old NAI guide, but try enabling Always discard next-to-last sigma.

Try a different VAE.

Hires fix usually fixes details like these too.

[lvcju]

Thank you for the answer, but I know what this guy is going through, it isn't related to operating the AI. Around the end of May something happened which I don't know, and the same generation data I was using and getting awesome results with little effort began to give results like you see in this post. I tried everything related to tweaks but none of the solutions I found addresses the problem, which confirms that isn't the operating of the AI. Also tried a lot of other things.

I made a post about it some weeks ago. #11214

[CaptDagon]

I think the culprit is the "Eta for k-diffusion samplers (noise multiplier; currently only applies to ancestral samplers (i.e. Euler a) and SDE samplers)" If I remember right originally it was set to 0 and was changed to 1 after an update, I remember playing with it back then and I had it set to 0.80 because DIMM gave me better results but If I set it to 1 it gave me some deformed details, I think eyes, mouths and other details were off, I need to some testing but I think thats it