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
The pinnacle of Perfection! Realistic yet beautiful.
My favorite version of this model to date!
FAQ
Comments (41)
This looks really cool, but you want 5k for a download?
Oh not at all. 5k is just to support the latest version and future developments.
You can freely download all versions that are 2 weeks older or less! <3
You can also click a button to get a notification when the current version leaves early access.
Just click on one of the other versions at the top bar and you'll get access to them all, even on the generator.
Thanks for checking out the model, I hope you will like it!
yea unfortunately i dont have the 5k for the new version, yea ill nest the next on from it and give it a try , the Preview images look amazing
@djacid No worries! Version 2.0 is greatly capable as well!
I think there's a pretty big step up to 2.2. It's based off of 1.2, where 1.2 is based off 1.0.
If you have any feedback on the model, do let me know please!
You mean 5k bucks:))) Thats pretty cool. You can make a base model from this.
@jeffreymarez765 5k bucks will not get you very far on a base model. They are ~usually 5 million bucks. And this is just the compute cost, assuming you get all the data for free.
5k bucks gets you a good unique checkpoint trained with lots of unique data. But it doesn't even get you a Pony or Illustrious which both use existing base models. Illustrious cost $180k.
I only take payment in buzzcks. Give 5k buzzks please.
@6tZ I mean a semi-base model. I have info only only about Bigasp, it was 3k, 6m dataset, joy caption. Its not good, but you can make one from 5k:)) Be total realistic, contain all celebs and amateurs, onlyfan babes and be on Illustrious.
see my next comment..
@Omikonz It's unfortunately not possible to micro-manage like that. I can lower it, but never raise it again after. Civit don't allow creators to manage it the way that you might want unfortunately.
@6tZ yah i goofed on the comment above. For those reading, All versions are available on the dropdown. Btw, I have found that the checkpoint is also great for a high-ress/variation fine-tune for anime into semi-realistic
Thank You.
It was a pleasure.
Amazing comments :)
If you have some more time later, I would love some on the cover images ^^
Amazing results V1 and V2. .. Thanks for this work. Looking forward to V3
v3 is being worked on! So far, I still prefer 2.2 though, so v3 is not ready for release yet. But I do have a spinoff model being tested.
Why your models cant work with usual samplers, schedulers. What that means? im courious. Are they overtrained?
It is a bit of an overfit merge this one. It can be hard to get it out of its comfort zone. Version 3 is improving this, but it lost a bit of luster compared to 2.2. Trying to figure out how I can have the cake and eat it too :)
I don't know why so I can't say unfortunately.
I wanted to see how the Perfection models are performing on the generator this first week of "pay to be in top 200", so I added them all, including the relevant version of Fucktastic.
I will not do this for all versions in the future. So if there's any version you like in particular, please consider contributing a buzz or two to keep it on the generator.
I'll keep supporting the top versions with buzz from early access of course!
<3 <3 <3
Not working mess
What's wrong?
Did you try it with the recommended settings?
Upscaling is strongly recommended.
@6tZ Yes, with recommended setting. Rendering just messy oversaturated/noisy blab. I upsaling with ultimate SD upscale in img2img cose i don't have that fast GPU (1080 ti 11GB).
@MafiapauCZE Could you share an image with me so I could take a look? I'd be curious to see what's going wrong here. I have a hard time getting bad outputs with it. Gut feeling is still settings.
I don't know the 1080, but my 3070 with 8gb has no problems to run the "hires fix", or equivalent in ComfyUI, which both are essentially img2img.
I haven't tried with Ultimate SD Upscale though.
@6tZ hires fix is just too slow for me so i'm using Ultimate SD Upscale because of that.
It doesnt work with usual sampler scheduler settings, thatswhy you get noisy images.
@ralpharnold715 I have no idea why it's not working. For me it plays just fine with any normal setting. I tried all normal samplers and they work fine. Some scheduler combinations might not work, or produces very bright outputs, but you can tune this down by lowering steps or CFG.
I tried the settings provided by OP, and they work fine for me.
The only thing that I haven't tried is "Tiled VAE". I don't have that option in my Forge.
So maybe that's the one causing the issues.
https://civitai.com/posts/15453720
This is with Euler A, automatic scheduler, same settings as OP provided.
This is with version 2.0 and 2.2 of the model.
@6tZ I only can talk about pony version, but its crying when it sees Karras:))
Would it be possible to make a version that uses clip skip 1 or 2? I need Compel to process longer prompts and better control the embeddings, but just using Compel requires that clip skip be set to 1, which makes the results not so good. If you make a version with Clip Skip 1 or 2, I will be able to use Compel and I don't think it would affect people who use your model through ComfyUI. I don't know if it is feasible or easy, but if it is possible, I would appreciate it 🥰
I honestly have no idea how to train or merge this into a model. Do you have any recommended reading on the subject?
@6tZ My hope was that you knew 🤣🤣🤣🤣🤣 We are all lost
@AishaAI I believe that clip is affected by any kind of training. But this means doing the training that resulted in the model from scratch. But this is mostly a merge. I have merged in LoRAs to fix specific things, but I'm not sure if it's possible to merge just a little bit to get a difference in Clip Skip behavior.
@6tZ As I said, I don't know anything about training, so I thought it would be something simple to do. I see a lot of models out there using clip skip 1 or 2, and they work fine when I use Compel. I don't know why Compel forces the use of the penultimate layer, which results in at least clip skip 1, but when a model like yours doesn't use clip skip, Compel starts to get bad, giving bad results. Your model works perfectly without Compel, since that way I can use clip skip 0, but without Compel I can't process long prompts because of the 77 token limit imposed by the default processing of text_encoder. I don't know how to get around this limit without Compel 😭
@AishaAI Hmm, I have no problems switching between different clip skip levels.
In ReForge (Forge/A1111), default behavior is Clip Skip 1, but I always run it at 2 and I have no problems actually.
These settings all work fine for me:
DPM++ 3M SDE with Karras scheduler, Clip skip 1/2 works fine, 30 steps, CFG 4.
What kind of problems are you having?
If you are getting blown out images, try lowering CFG. I went down to CFG 3 for Clip Skip 2 to look better.
All the following samplers worked fine for me:
Euler a
DPM++ SDE
DPM++ 2M
DPM++ 2M SDE Heun
DPM++ 2S a
DPM++ 3M SDE (try this one with Karras scheduler)
@6tZ I have the same problem in different schedulers, which is the drop in accuracy and deformation of some things.
Here's a small example:
Prompt:"1girl, blonde, blue eyes, smile, hand on waist"
If I use clip_skip=0 without Compel, 8 out of 10 generated images will show perfect hands, which is a good result.
When using Compel + clip_skip=0 (which will have the behavior of clip_skip=1), 7 out of 10 generated images will show bad/partially deformed hands, which is a bad result. Increasing clip_skip does not improve the result.
Without Compel, when using clip_skip=1, the results start to show deformations, just like Compel does, so my conclusion was that I lost accuracy whenever I used clip_skip>0.
I'm not using any correction tool like upscaler or detailer. Compel is a great tool, but in this case, not using it is worth it due to the quality of the model. The only real problem is that sometimes the embeddings generated by the default text encoder are not good, causing some parts of the prompt to not be followed, something that Compel solves very well. In any case, this is a problem with the way I use it, and not a flaw in the model's training. It really is very good 🥰
@AishaAI Would you mind showing me to Compel? I tried searching for it but googling doesn't give me any credible results on a tool called Compel.
@6tZ Compel is used with diffusers https://github.com/damian0815/compel/blob/main/doc/syntax.md
You surely know the Compel syntax
@AishaAI Hmm, I really have no idea what this is or how to support it :D
I only saw this in Easydiffusion(++ instead of number). I dont know, if this method is better at all.
@ralpharnold715 Well, it probably uses Compel, as that is part of the syntax
@AishaAI And is it better? I switched from it, because of XL.
@ralpharnold715 Compel is very good, it is a tool that helps to manage and choose embeddings better. The only problem with it is that it actually forces the use of clip skip 1 and some few models perform better in the last layer of the clip (or clip skip 0). Compel is more of a tool to obtain precision, it will improve or worsen depending on the use and compatibility. In general, it helps a lot, but in my case using this model, Compel is hindering the quality because of clip skip 1. And Compel works on any basemodel, because it is an embedding manager. So regardless of Pony, Noob, Illustrious or XL, it works
lindo
Well I gotta admit this is Perfection!