DreamShaper - V∞!

Seriously. I have been trying for 3 hours straight, no fuckin' luck.

@lykon https://imgur.com/eAn2F5p I have been getting this image when trying to recreate the one of the girl from behind looking at the ruins

@lykon Did everything, ended up with this result https://i.imgur.com/9tLycZs.png. Also tried clip skip 1. Im not using any parameters (xformers, medvram, lowvram, etc...) but i dont think the difference should be that big. Tried the one with baked vae and also without and got the exact same result (expected)

@lykon Thank you very much! with 0.35 i could almost get the exact same image as the original. The embeding "bad-image-v2-39000" (https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt) was extremely important and then the "by bad-artist" (https://huggingface.co/NiXXerHATTER59/bad-artist) embedding adjusted the image a tiny bit to almost be there. Here is the result: https://i.imgur.com/vJzVucr.png. I dont know what else might be the difference between this image and the original.

"I had ENSD: 31337 for basically all of them"

You need to go to Settings>Show All> Find "Eta noise seed delta" and change "0" to "31337"

@lykon theres no mention 'mksks style' in prompt, only 'anime screencap style', which i downloaded and used, but still not even close. I used the old 3.3 version (without clipfix and baked vae)

Im new to civitai and all the model stuff, so forgive me. I just saw this question asked and I was curious- What is mksk? How do we see the prompts? And what is "Lora"?

@lykon the 'anime screencap' inside the prompt is just a placebo then?

i still don't understand why i'm not getting same output 😬. I downloaded the loras, tried them, but still not same result

@Lykon thank you for responding, I used wanostyle-20 and i've got the same result as yours. Thank you so much for all your work!

like what? can you upload some samples?

ENSD set to 31337.

It's in setting: ENSD is short of eta noise seed delta.

@Lykon 
https://civitai.com/gallery/39708?modelId=4384&modelVersionId=5213&infinite=false&returnUrl=%2Fmodels%2F4384

does this link work?

@Lykon By changed I mean the seed changed from 105259063 to 105259061 for the two images

The two seeds are the only difference in the 'quick test after fix' photos for version 3.32 bakedvae (clip fix). So why can't I replicate the same image?

@Lykon They do look good, yes. I did copy the same generation data including the seed and tried sliding clip skip to 2 and 1 but both didn't exactly replicate. I think I just want to make sure I'm getting the most out of this model and thought that if I replicated the example images exactly then I've got a great starting point.

@Lykon Waaait I got it! https://i.imgur.com/5HmSAwR.png Clip skip: 1, Seed: 105259063, Batch size: 2. The batch size threw me off, I only did size of 1. Anyway, I feel better knowing that my set up is not abnormal.

i had a few successes, with (dragon:1.2) girl and similar, but without a TI/LORA it will always be up to luck what happens. keeping 9:16 aspect ratio, like 480x872 will increase the chance of a blending, while 16x9 will almost surely try giving a girl and a dragon.

ok, makes sense.....does anyone know a good one for this type? maybe not specific for dragons but a more broader one? havent found anything so far

@Lykon pls do. if you need help finding data, then reach out^^

@Lykon Ayo, same here! Im moving house but I can see engineers are installing 900mbps internet :(, we got like 5mbps. It is very upsetting

@Lykon but all the lora that I downloaded is .safetensors and this mksk file is .pt

@Lykon 

Yes .

Unfortunately it has stopped drawing.

So I tried making Lora by copying his method but I fail every time.

If you do that, I will be very grateful to you. And thanks for the reply ❤️

great artist

@Lykon Thanks for answering. As for the faces, any tip to improve it? Most of the time seems one of the eyes is off. I'm really new with AI art so i don't know if im doing something wrong.

Use full res inpainting on faces with full body shots, i have a guide here: https://civitai.com/models/1366?modal=commentThread&commentId=2828

is simple to understand. you have a limited amount of data points that go into a image. so if you have 100 points and a close up portrait, 80 points may go to create the face. so it looks god. but if you create a full body shot, to the face they may go 5 points. I made these numbers up in reality the real data is way higher. But you get the idea. So to fix this you need to use the full res for the faces inpainting technique, which will draw the image again only on the masked area but this time giving it the entire 100 points to work with

@dilectiogames perfect explanation.

@Lykon Thank you for your work, will be paying attention =)

You can make an inpainting version of any 1.5 based custom models yourself in webui by merging it with 1.5 SD inpainting and then removing 1.5 SD

oh wow, i had no idea thanks for the info and the inpainting model.