[Stable Diffusion] Prompt Sharing and Learning Thread

[devilkkw]

Sepheyer , never mind these ignoramii These pictures were really nice: I love the hilly mountainous ones! This is for the others, so they recover from not seeing titties View attachment 2591018

Is all good result, except, to me, the skin look like photoshp processed, especially in face. are you using LORA for face? Seem all image have same face, is it's because you triggered a name? (sorry i don't recognize it).

 

[osanaiko]

Is all good result, except, to me, the skin look like photoshp processed, especially in face. are you using LORA for face? Seem all image have same face, is it's because you triggered a name? (sorry i don't recognize it).

I'd say it is a custom Textual Inversion to get a consistent "character", perhaps following the guide that Mr Fox posted a few page back.

 

[sharlotte]

Yes that's what it is, created my own TI to get the same face throughout. That's the Sharloth37v2 you see in the prompts. The skin though is generated via the SD, however I tend to always use the same 'style' so that may be it: 'RAW photo, a photo of Sharloth37v2, (high detailed skin_1.2), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT' I may vary the 'lighting' bit but the rest for the past few weeks has always been a constant.
Depending on the model used, the face may or may not come out nicely (for me that is) and I usually stick the created image in 'inpaint' and create multiple versions with different expressions, blurring the face only (as per Mr Fox advice a few weeks ago, I tend to generate quite a few iterations of an image and just keep the 'best' which I then work on). Never had to use photoshop so far (but then it's only 3-4 months since starting to use SD or any of what we have been discussing here).

 

[PandaRepublic]

I asked this question like a month ago, however i'm still having issues with some of the LoRA's I have made and I want to know what causes this problem. Even lowering the weight of the LoRA the problem still persists. It seems like it loads the LoRA pretty well then it shits on it halfway through the render.

preview image when rendering:



final image:
​

 

[KingBel]

I asked this question like a month ago, however i'm still having issues with some of the LoRA's I have made and I want to know what causes this problem. Even lowering the weight of the LoRA the problem still persists. It seems like it loads the LoRA pretty well then it shits on it halfway through the render.

preview image when rendering:



final image:
​

We would probably need to know more regarding your settings and config in order to assist, but one thing you should check is to make sure that you are not using Face Restore. It will remove any face details and smooth out your final result, which is often not what you want when you are trying to recreate a specific face.

Also I have found (with embeddings in any case), the result will be different from one checkpoint to another.

 

[Elefy]

Face Restore. It will remove any face details and smooth out

 

[Mr-Fox]

I asked this question like a month ago, however i'm still having issues with some of the LoRA's I have made and I want to know what causes this problem. Even lowering the weight of the LoRA the problem still persists. It seems like it loads the LoRA pretty well then it shits on it halfway through the render.

preview image when rendering:
View attachment 2592050


final image:
View attachment 2592051​

I agree with Kingbell, try without restore face. I use postprocessing most of the time, sometimes in combination with face restoration but it's not always the best. Denoising could be it, if you are using hiresfix. I don't think it has to do with the Lora itself but rather how you generate the image with SD. In regards to training Lora's. Slow training is the best, both from what other's have reported and also my own experience. If you used generating sample images during the training to see how it was going it's normal to have the images not be very pretty. But when you generate images with SD it should look great.
I know that some tutorials say that it should be possible to use with 1.0 weight strength, it seam a bit idealistic to me.
My own Lora works great up to 0.8 . What is absolutely the most important is the quality of the sample images that is being used for training and how much effort is put into preparing them. Then the settings and on wich checkpoint you are training it on. I had good result with elegance, I tried of course with clarity and some others but it wasn't good for this.

This is the absolute best info source for training Lora's that I have found.

You must be registered to see the links

 

[sharlotte]

I never mentioned but the reason you don't recognize the 'woman' is that she does not exist: I followed the thread posted by KingBel in #1,289 and created a new persona, my very own Sharloth

Here using similar settings, what the AI does using high res with similar prompt but no TI:


I think the last one does not really look like a Brigitte Bardot as a 19yo. But it still is a lovely image.

 

[deltami]

anyone may suggeat a good checkpt/TI/lora for vellus hair in really close up skin?

 

[Mr-Fox]

After some searching I found

You must be registered to see the links

(checkpoint). It looks promising for this purpose. In the sample images you can see skin defects, wrinkles and peachfuzz (vellus hair), even looks like it has skin pores.

You must be registered to see the links

 

[Elefy]

skin indentation, skin pores, textured skin

Looks like work in every model

 

[KingBel]

skin indentation, skin pores, textured skin

Looks like work in every model

I may have mentioned this before.. but there is an upscaler called SkinDiffDetails which is excellent at enhancing and maintaining skin details and imperfections. On this wiki (under the Skin category):

You must be registered to see the links

 

[Elefy]

and I overdo it hahaha

 

[Kaseijin]

It's impossible without the meta data. We can only guess and speculate. Or you can message the poster and ask pretty please with sugar on top.. Insta account is dead fyi.

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

Nice work!

Check this user out too

You must be registered to see the links

You must be registered to see the links

 

[Mr-Fox]

I may have mentioned this before.. but there is an upscaler called SkinDiffDetails which is excellent at enhancing and maintaining skin details and imperfections. On this wiki (under the Skin category):

You must be registered to see the links

NMKD has also face focused upscaler and other's

You must be registered to see the links

.

 

[Mr-Fox]

View attachment 2592966

and I overdo it hahaha

If you would tone it down a little it could be great. It has great potential.

 

[Jimwalrus]

If you would tone it down a little it could be great. It has great potential.

It most certainly does - here's my attempt at generating it in one go, with a few tweaks to the settings* and CyberRealistic2.0 as the checkpoint.

*Lower res for a start! I've got 12GB of vRAM and that was still falling over.

 

[Nitwitty]

Has anybody checked out the Hyper V1 checkpoint yet? It's excellent for photorealistic models. You can find it on Civitai. Anyways, I'd like to know where we are on creating consistent models if this technology is going to replace DAZ for making visual novels. Also, how do we address the issue of hands and feet. For me, I'm more concerned about feet and soles (heh heh).

 

[Nitwitty]

Trying a new model Cyberrealistic on Civitai (

You must be registered to see the links

) and working fairly well so far:
View attachment 2589479 View attachment 2589480 View attachment 2589481 View attachment 2589482

Do me a favor, I want you to check out two other models - EdgeofRealism and HyperV1. Both are great at creating photo-realism especially HyperV1. I like cyberrealistic though.

 

[Sharinel]

Here's the same prompt within HyperV1. Also shows the difference in upscaling. Both these have the same seed, but the 1st one is upscaled 3 times and 2nd one is slightly bigger original but upscaled twice to give the same output