Welcome to Perfection Realistic!
This is an Illustrious XL merge with a bunch of hand-selected checkpoints and LoRAs to improve the base model's realistic capabilites.
And wow, does it deliver!
Also available on mage.space.
It's focus is on creating high quality, photorealistic, but slightly exaggerated images with a focus on women.
Do not use the Normal scheduler. Use Simple or Beta!
Use euler_a, dpmpp_2m_sde or dpmpp_3m_sde as your sampler.
To use the faster low-step settings, you need to use the DMD2 LoRA. This allows you to use 8 or 16 steps and generate very fast. Remember to generate in 2 passes, either using "Hires fix" in Forge/A1111, or by creating and chaining a second sampler node in ComfyUI. A sample workflow if included in the sample images of the model. Use this as a reference.
Each sample/preview image contains the used workflow. Here's a quick article with simpler more beginner-friendly workflow. This is a recommended starting point.
Recommended Settings
Base Generation
Steps: 24
CFG-scale: 4
Sampler: DPM++ 3M SDE
Scheduler: Simple (beta upscale scheduler)
Upscale/Refine
Steps: 8
CFG-scale: 6
Sampler: DPM++ SDE (dpmpp_sde)
Scheduler: Beta
Denoise: 0.35
Description
FAQ
Comments (35)
I like 3.1 but early access model every week feels kinda grify ngl
It wouldn't be that bad if the buy-in for early access was only in the hundreds each release, you know? Not, what is it now? 6,000 buzz? That's absurd for each release version lol.
I hear you. To be fair, it's only every second week, as this is how long the early access lasts ;)
But I understand that's not what you mean.
I prefer to see it this way instead:
Every other week I give everyone an all-new model I've spent days testing and refining, for free.
If you see something that looks exceptional, you could consider donating some buzz for it, either to unlock it, or just to show appreciation. I have tried with different amounts, and sometimes I have a lower cost on a model to let everyone access something new and cheap or for free. I did this with Perfection 2.5D the other week. One free version, and one a bit cheaper.
I'm still generating images with version 2.2 because I find this version to be perfect. There's nothing wrong with it just because there are newer versions, right?
But I must also be allowed to "push" newer version to the top, to make them more visible. This seems only fair.
I hope it explains things from my point of view.
I set the price to 5555, but I guess it rounds up on the visible price, so I guess this was a bad idea. I'll lower it to 5000 instead.
@6tZ I'll buy it just for the sincere reply
@puz5066 My evil plan worked! I'll convert all the doubters with kindness until they are all my buzz-slaves! Mwahaha!
Thanks friend!
One final note on the topic:
A higher version of the model does not mean higher quality.
Of course it's all iterations and "improvements", but some times it's a step back, or a sidestep in a different direction. Based on what you want.
I like version 2.2 because it's highly cinematic. So much that I branched it onto the Perfection Cinematic line of models. Some might light a more raw style, or realistic low quality images, because they are more realistic.
To summarize, don't feel the "fear of missing out" just because you're not on the latest model. If a checkpoint is interesting, I usually try all versions of it to see which one I like the best.
Of course, I don't have time to try every single version of every checkpoint, so I usually go with the free ones, and then give donations or buy something just to support when I see something extraordinary or interesting/different that I want to support. That's how I do it, but everyone should do what they feel of course.
Thanks for coming to my TED Talk.
@6tZ Gotcha. What is your general approach w/ the merges? Is a new version of a checkpoint based on the previous one? Like is v3.2 mostly the same checkpoints, with some tweaks to weights, etc? Do you merge LoRAs in too?
@puz5066 Good Q. I have no set method. I do some additive and some subtractive merging, as well as normal % blend merging with checkpoints. And I do LoRAs as well based on styles I'm trying to implement as a core look, or add functionality or poses.
As an example, I noticed that earlier versions were quite incapable of doing any kind of ahegao face, so I softly added this by merging in LoRAs of the pose/expression, while trying to be careful to not go too far, to make every face like this. Delicate balance. It's still not perfect, but improved.
If there's something specific that I'm looking for, I usually train one or two quick LoRAs and merge them in as well. Skin details, aesthetics, diversity, stuff like this.
You can also iterate a model by training the checkpoint itself, but it's less flexible compared to being able to blend in a value with a LoRA, and it's also more annoying (GPU intensive), to train. I prefer to throw some images at the Civit trainer while I keep working on the model, and then an hour later I have results I test and merge into the model.
This is the process I mostly follow.
Again version 3.2 brilliant. I compare it with all available photorealistic models and the output of your model is brilliant... at least for the images I create. You really know what you are doing and I am very pleased with your work. So again, this version deserves my support in the form of 5000 credits because it really does a lot and looks great.
Thank you kindly.
I mean absolutely no pressure to keep up with every version, but I genuinely appreciate your support.
Maybe I should be setting up a patreon for repeat supporters to get a better deal?
@6tZ That's totally cool. When I'm happy, I don't regret the gift 😎👌
amazing :)
My goto model. For each idea I have I try some different models but whatever it is I'm trying to do, I find this one the best every time! I'm often using CFG 5.0, Euler a, 25 steps, with the same settings for high-res but only 12 steps.
Can't wait until I can use 3.2 in the website!
This model is excellent
So I paid 5000 for early access to 3.1 and now I can't use it anymore? I have to pay another 5000 for 3.2? :|
Is that true? Even after dropping 5,000 buzzes to get the 3.1 model, you can't use it again because of the 3.2?
Shoot, yeah that sucks :(
3.1 is not on the generator any more due to how Civit works.
Each model needs to pay to be on the generator each week. This week I paid to have 3.2 on the generator.
But anyone can chip in and pay for older models to be available as well. Usually it requires between 7000-10000 buzz to be on the generator for a week. So the model needs to be quite heavily used to reimburse that investment and even go +- 0.
I put in some other versions this week. But if the interest in 3.1 is strong and I get some help to cover the cost, say 50% of it, I will put 3.1 back on for next week.
Feel free to complain to civit staff about the system :)
Note:
This is only for using it on the generator.
Of course any early-accessed model can be downloaded freely.
I like it, but it seems to really struggle with hands, majority of the time only 3 fingers - far worse than any other illustrious model I've tried.
That seems odd. I occasionally get 4 fingers. What settings are you generating with?
@6tZ I've had another play. Without any LoRA it is behaving itself (24 steps, DPM++ 2M SDE Karras, CFG 4) So looks like I'll have to be careful with the loras! It was the 3.1 version by the way. Thanks for your work!
@jimgrainge758 Ah yeah, it does not play amazingly with other LoRAs.
It tends to fight for control, which can contribute to things like this. Lowering the strength tends to help it a bit.
Hi, if you mean 3 fingers and a thumb, ET style then yes i have notice this as well. But hey I will take ET style hands any day over the more common problem people talk about with hands being the 5 fingers and a thumb which they call 6 fingers.
I love this model <3
Its my favorite!
Why can't we use this model directly on the site anymore?
Do you read the news? Civit is going byebye soon with all their censorship. OP is probably just protecting his model from being deleted.
@punkbuzter340
No, I usually don't read them because you see a lot of bullshit.
Ah no, CivitAI has been for the last couple of months on an "auctioning" schema for their 200 models available for onsite generations.
Basically, I have to pay them buzz to have my model be on the generator. And unfortunately, this week I got outbid with each of my bids, as I couldn't stay up and "guard" my bids.
Now with them not being able to accept card payments for the time being, model donations and buzz income with generations became less than half of what it used to, so it makes even less sense for me to pay $10 for each model to be on the generator each week.
You will notice on a model that is NOT on the generator, that there is a small teal "Bid"-button at the top left corner of the side panel. If you click it, you will be able to contribute buzz towards your favorite model to be on the generator. If the model does not get selected, you get your donation back, so it's never lost, only spent if the model you supported placed in the top 200.
So, if people really want to see the models on the generator, and support the models with bids here, I will of course also chip in.
So please consider contributing a small amount of buzz for the models you wish to see on the generator.
@6tZ this blows. Your model is easily the best one I've used. Hope they fix the cc thing soon.
@MFKall It is somewhat fair I think, the most popular models should be on the generator. I kind of wish there was a stronger rotation though. Like some slots are for staff picks or community voting, but not with buzz. And it could only be on new models, or previously not featured models, or something like this.
Anyway, anything can be put on there, as long as you pay them...
Hi - This has become my favorite model and I donated to your early release campaign. Will you be making it available for use on the site rather than just for download?
Hi Henry. Thanks for the kind words.
CivitAI has for the last couple of months been on an "auctioning" schema for their 200 models available for onsite generations.
Basically, I have to pay them buzz to have my model be on the generator. And unfortunately, this week I got outbid with each of my bids, as I couldn't stay up and "guard" my bids.
Now with them not being able to accept card payments for the time being, model donations and buzz income with generations became less than half of what it used to, so it makes even less sense for me to pay $10 for each model to be on the generator each week.
You will notice on a model that is NOT on the generator, that there is a small teal "Bid"-button at the top left corner of the side panel. If you click it, you will be able to contribute buzz towards your favorite model to be on the generator. If the model does not get selected, you get your donation back, so it's never lost, only spent if the model you supported placed in the top 200.
So, if people really want to see the models on the generator and add buzz to the Bid-section, I will of course contribute myself as well. If 10 people using the model would add 1000 buzz each ($1), the model would be in the loop for another week.
I hope this helps and explains it.
I dont understand the hype about this merges.
It barely knows anything, what illustrious is. Sex poses are laughable, just put them into XL category.
Yeah I can agree with you in terms of similarity to Illustrious and knowing what it knows about anime characters. It barely knows any of them any more. It's something I'm trying to blend back in there with future versions. But trying to keep this aesthetic.
@6tZ Its not about anime characters, but sex poses doesnt work in booru way, its not usable. Booru sex prompting is the easiest in the world, now thats not working too.
@shlomylindan691 right because a cloud tag makes sense. lol wtf man. Purely for natural language. purely purely: I don't know man Flux is might have the edge there LOL. what part of: 1girl ((()))))))))), guy, boy band, 33 wives, no ptsd, nightmare fuel, or even the kitchen sink. that doesn't make sense compare to: Elegant Theater La Grotesque, gouche style painting with hints of cubism. The subject is a college age looking curvy, woman bouncing like a manic spring on some dudes dick.
what version is thios critique targeted 3.0 or which one?



















