@epinikion you bet. love these models!

@epinikion Then answer me, I'm waiting)

@Luciferix one has more Elon Musk in its training data

@epinikion 64 of what :D

I've been using these with all the Epic Loras and embeddings together, the realism is real! Only one model has currently rivaled them, but I won't promote the enemy here hehe

@epinikion No no need to be sorry, crazy in a good way, you give us choices, I think that is the nature of all models none are perfect, none do all and all do something great so we end up with lots, then comes matching the prompt that model likes vs the negatives they like.

Bake away. :)

This is not a SDXL model, and i assume you installed SDXL only settings on your web-ui

The funny

Good point. Any "updates" should include a description of what is new to the model.

@FrozenGunner check 'about this version' and author can't decide which one is better haha

@sgmgong28901119 

Ahha, arms, hands and legs.

@sgmgong28901119 yeah

Thanks

Something like x = +1000 realistic girls added to the inteliggence
x plus = + 100 realistic girls + 2000 epochs/generations/work
x plus plus = + 100 realistic girls + 2000 epochs/generations/work

@schsch 

OK, OK, don't rub it in. I'm pretty new to this and didn't notice there was a "About the version" thingy on the side 😁.

more cow bell

Well, in some tests, people are saying X is more stable than X++.

I am truly assured it was a coincidence. Nobody here cares about differences, we all are human beings.

@epinikion I am using 3M SDE. I am not using a VAE as I thought it was already embedded - isn't it? Should I be using a VAE?

@epinikion I have not! The "++" naming was too good not to test first, like "but this dial goes to 11"

@epinikion X+ understand prompts better than X++ at CFG scale 5, similar to x

@epinikion I've found that X+ is more stable than X++, but X is the most stable of all.

Alle three are great in skin and faces .. but also some cons .. the plus (X+) is for my prompts better in anatomy. All three versions are struggling in landscape (more width px than height px) prompts. Thanx for sharing and creating those wonderful models.

@joge25 Yes, only SDXL works in higher resolution

create lora/embedding/dreambooth of person and use created file for img2img with controlnet openpose (most control)

use same seed on img2img and slightly change description (inconsistent)

use inpaint to change worn clothes or part of environment (limit to existing photo)

use tools that swap face and apply after txt2img photo generated

Try changing your prompt slightly. Say you show redhead woman standing in doorway. Regenerate that picture but change the prompt slightly like "redhead woman sitting in cafe with sweater and hat", or "running in park with sports bra and yoga pants" or standing in office with blue blazer, cream blouse, and long black skirt", etc. Just experiment with different scenarios.

X was a refinement on v4 (which if you haven't tried v4, you should give it a shot) to get even better physical anatomy, but the author was afraid that they ruined its ability to produce good hands in the process. So X Plus and X Plus Plus are a couple extra merges that were both attempts to make hands look better, but they couldn't decide which was better and released them both.

I really haven't noticed any major hand problems with X. For that matter, I think v4 still generates great images in terms of anatomy.

@Invisidude thanks. You would wonder why the person would not just say that in the description.

@KINGLIFER yeah, it's kind of buried in the version notes for each version.

@Invisidude Thanks for the info man! appreciate your time.

Had the same thoughts recently, man

For me, it's V3.1 get what you prompt. Absolutely stomped by how great it is. I'll have to test X++ to really see if I'm going to stick with it.

That is the common issue of burnt mixes. Despite the good quality it outputs. Results are burnt in a certain area.

The way I usually deal with that is by using ADetailer, because it detects and inpaints a new face with varying degrees of difference. You can set a specific prompt for different facial features, different from your main one. Then you get many different faces and they are much better quality too! Win win!

Controlnet is also another great way to get good variety of faces. I usually use Depth, but lately have been using lineart, and have been getting pretty good results.

Use random names instead of woman or girl. Search for the native female names for the ethnicity you are looking for on Google and try.

@Nobodybutmyself Mind breaking down a bit how line art could vary the face?

There are other ways to vary the face. Lora is probably not the answer you want but I'd suggest using Lora or a common name/famous name to add to variation. Now with the lora of your choice, you don't want it to completely dominate the render, but to just add slight detailing.

So something I learned from researching mixing faces was to follow this formula. Ensure you space it exactly the same.
[woman: dove cameron: 0.8], blah, blah, blah.

So for the first 80% you'll get your standard woman face from the model itself then it will add dove cameron's face for the last 20% of the render as a detail. You can even add more faces with something like this...

woman, [latina: viking: 0.5: dove cameron: 0.8], causal outfit, fur shoes, blah, blah, blah,

You get your gender start, with your race, then another race typing start at 50% of the render then at 80% of the render, dove cameron's face is overlaid.

But you need to follow the formula within the brackets, otherwise it won't render the face like you want it. You can fiddle with percentages how you want. Plus, you will get a consistent face. You can even add Adetailer to the mix to get better detailing.

But this is probably the best method of getting a truly unique render. Obviously you can use other names BTW.

@pekonaxdd So, just to elaborate on my previous comment, I've decided to render a few examples for you. You can use the prompts and adjust it to your taste. But you will find that the blend feature is extremely powerful, almost to a perfection. You can even create an entire line of ancestry if you wanted.

My examples will be found under 3.1GetWhatYouPrompt, as that's the version I use. Please, please, look at them. I want you to really get the gist of what I'm trying to bring out. Hope my little display and comments will help you in your search.

@superjosh250 duuuuuuuude this works so well!!! Thank you very much :)

@automatas Glad it works for you! I didn't use Lora for the images I posted, but I highly recommended doing so for the greatest likeness of a celebrity. However, its not necessary as you probably have noticed. Just using fairly well known names, that the model already knows workings for blending.

I'll post a lora blend soon just to show how powerful this mixing method is.

@automatas Glad it works for you! I didn't use Lora for the images I posted, but I highly recommended doing so for the greatest likeness of a celebrity. However, its not necessary as you probably have noticed. Just using fairly well known names, that the model already knows workings for blending.

I'll post a lora blend soon just to show how powerful this mixing method is.

I had the same issues. The easy way to do this is to just use a specific name of a famous/semi-famous person in your prompt, and the AI will put the face into the image. That's partly why I focused on just putting images of faces because I knew others here could experiment with them in their own AI art. I focused mainly on Caucasian blondes, Black women, and brunette women, as most Caucasian brunette women could be used for Latina women as well. I've only tried one Arab woman, but I'd like to work on more Arab/Palestinian women as well as Persian and South Asian women. Unfortunately, I'm not as tuned into Bollywood as some others here are, as I only know Priyanka Chopra and Sunny Leone.

X was a refinement on v4 (which if you haven't tried v4, you should give it a shot) to get even better physical anatomy, but the author was afraid that they ruined its ability to produce good hands in the process. So X Plus and X Plus Plus are a couple extra merges that were both attempts to make hands look better, but they couldn't decide which was better and released them both.

@epinikion whats other files about? X-plus? X-plus-plus?

I agree. Which version do you prefer? X, X+, X++, etc?

@civitai_dane Im using X,it's the best version.

@epinikion Yeah and i get to conclusion after many hours of testing that InvokeAI is now with version V3.3 much better then Automatic1111,upscaling is smoother and much faster with lower memory consumption,and in SDXL support InvokeAI is also much faster.

you can upscale with natural high resolution Fix enabled to 4k with only 8gb of Vram on Automatic you would need RTX 4080 with 16gb of Vram which is absolutelly crazy.

From now on InvokeAI is my daily driver,i forget about automatic lol.