Cyberfix - Cyberfix/perpendicular V2
This is similar to past cyberfix-perp/cyberfix version, but using cyber realistic V4 instead of 3.1.
Separated in a different post since the past one was too cluttered, and also to separate from EPS and VPred.
Past versions on: https://civarchive.com/models/1001795
Original description
Cyberfix perpendicular
This one is thanks to DoctorShogun, more info and model can be found here https://huggingface.co/Doctor-Shotgun/NoobAI-XL-Merges/tree/main, as he made this merge. You can checkout his civitai and loras here https://civarchive.com/user/DocShotgun
Instead of doing the normal recipe for cyberfix, it uses a perpendicular merge into it using mecha.
Let us know how it goes!
Continuation of Cyberfix
As I posted the models on Huggingface, some people have asked me to upload the models here to easier usability, which of course I have no problems.
Please if you can add images, it would be appreciated!
Models with cyberfix now (ordered by model release date):
noobai-XL-Vpred-0.5-cyberfix
noobai-XL-Vpred-0.6-cyberfix
noobai-XL-Vpred-0.65-cyberfix
noobai-XL-Vpred-0.65S-cyberfix
noobai-XL-Vpred-0.75-cyberfix
noobai-XL-Vpred-0.75S-cyberfix
noobai-XL-Vpred-0.90r-cyberfix
noobai-XL-Vpred-1.0-cyberfix
You can also find them on Hugging Face on my profile: https://huggingface.co/Panchovix
Original model: https://civarchive.com/models/913998/noobaicyberfix?modelVersionId=1022962
NoobaiCyberFix
Purpose of this merge
The purpose of this model is to fix the anatomy problems of model NoobAI, while keeping 100% of its knowledge and style. NoobAI-Epsilon075 is a great model, it knows almost all characters and styles from Danbooru, but it has some anatomy issues and this merge completely solves this problem. If you use NoobAI-Epsilon075 for SFW purposes, you will hardly notice any difference between this merge and the original model, because the goal was to keep all the knowledge of NoobAI-Epsilon075, and visually this model is almost indistinguishable from NoobAI-Epsilon075. But if you are a man of culture and like Pony models, you will definitely like this one, because this merge gives the model a very good knowledge of anatomy, almost at Pony level
Recommended negative prompt
lowres, (worst quality, low quality, bad anatomy, bad hands:1.3), abstract, signatureRecipe of the merge
1 Create temporary model
TempModel = NoobAI-Epsilon075 + cyberrealisticXL_v4 - stable-diffusion-xl-base-1.0
2 Remove CLIP from TempModel and replace it by CLIP from NoobAI-Epsilon075
I did this step in ComfyUI.
Here is the workflow
https://civarchive.com/models/991574?modelVersionId=1110897
Logic behind this merge
cyberrealisticXL_v31 is a general purpose model without any specific style, just like stable-diffusion-xl-base-1.0, and so if you replace the stable-diffusion-xl-base-1.0 weights with cyberrealisticXL_v31 weights in any model, it shouldn't change the model's style and its knowledge much. But at the same time, cyberrealisticXL_v31 has much better anatomy knowledge than sd-xl-1.0, and because of this, it should greatly improve the anatomy knowledge of any model, and in the case of NoobaiCyberFix, it worked exactly as I expected. In theory, you can use this recipe to improve the anatomy knowledge of any model that needs it
Big thanks to the creators of these models!
Description
FAQ
Comments (44)
v2?have new optimizations！！
Mostly it changes cyber realistic from an older version to a newer one.
IMO, I like the newer one more, but just an opinion haha.
Pancho the goat. Only thing making Noob even remotely usable
[Solved] What realistic and base models to use fixing a vpred model, please?
I tried with another realistic model for noob-vpred 65s, but it gave noise output when generating images.
EDIT - [Never mind, turns out the output models weren't the problem, my workflow couldn't handle the generated models.
Which is weird because the models downloaded from here worked just fine in the same workflow.]
you will basically need always
sd_xl_base1.0
a realistic model (for this case, cyber realistic v4)
a base noob model
I have mostly tested with cyber realistic, I guess if there was a realistic model with vpred, results would be better.
@aflab Yeah, that works for epsilon models I guess, as that's given in the workflow provided as well.
But the trying to fix a vpred model with that gave just noise. So, I guess some of the other models must change or some other node goes in between to take care of that, may be.
I was hoping @Panchovix would chime in.
@aflab Turns out, the generated models weren't the problem. I've edited my original comment accordingly.
Nice you managed to find out!
I have a severe case of skill issue. The model behaves very well... Until it makes hands or be used as upscaler. Whatever Noob finetine I use - it fails at hands (they look very distorted or as in 20 finger fails in case of upscale) and upscale (it always look pixelated, noisy etc.), the image has these oily spots from latent noise as if it wasn't denoised properly, it remains with whatever combination of sampler and scheduler. Which is not a case for mixes or merges of Noob v_preds. What am I doing wrong?
Did you found the solution? I'm having these issues with this checkpoint and also don’t know what Sampler should I use 💀
I still think that mentioning "Illustrious" model type is wrong.
Illustrious is not the same as NoobAI-XL. LoRAs from IL doesn't work properly on NoobAI-XL and vice versa.
NoobAI-XL is much bigger finetune than Illustrious.
Especially NoobAI-XL V-pred has completely different prediction type (V-prediction instead of Epsilon which is used by Illustrious).
I hope they'll get own model type "NoobAI".
Well the eps version of noobai works with normal illustrious stuff. Imo civitai needs a new category for NoobAI V-pred.
I think the issue is there no better category to put it out.
If you use just SDXL, people that filter for illu/noob loras and checkpoints won't find this model.
And if you use Ilustrious, it will show also models that aren't noob.
The best solution as you say, would have a "NoobAI" category.
@Panchovix It is sad that pony which has only 2.5 million art finetune and Illustrious which has 7.5 are labeled but the biggest finetune which is better than NAI3 is not labeled 😥
I agree that as long as there is no separate NoobAI category, IllustriousXL is the best one to choose from. NoobAI is a great model but it is still based on training from IllustriousXL. This is clearly stated by the creators of NoobAI on their HuggingFace pages and it therefore inherits the license of IllustriousXL. And LoRA compatibility does not define if models are essentially different.
@lizardon1024 There is another problem. If I change base model of my loras on Illustrious I won't be able to put in recommended checkpoints NoobAI-XL (Eps and V-pred). I want to put in recommended resources original model and cyberfix but I can't. So since original model is labeled as SDXL 1.0 I must to label loras as SDXL 1.0 too and I can only mention in description that the best model to use is cyberfix.
P.S. Actually cyberfix has 1 issues in comparison with NoobAI V-pred - too bright lighting on character even with tags "shadow, shaded, night, dark".
@Enigmata So not only would CivitAI need NoobAI as selectable base model, but also v-pred and cyberfix. We can just put it on the list of all things that should be changed on CivitAI. With only the choice between SDXL 1.0 and Illustrious, NoobAI should have chosen Illustrious instead of SDXL 1.0 as base model. Marking it as general SDXL only increases the confusion...
@lizardon1024 I changed all my noobai epsilon loras to illustrious but saved V-pred as SDXL 1.0. I think it is not fully right but if people looking with IL filter for noobai loras they at least can see noobai lora.
For now, I've changed the type to SDXL 1.0
The civitai generator does not detect it, can you upload it?
What do you mean civitai generator? Sorry, I don't quite know how to use the civitai eco system
@Panchovix Civitai won't let me run your checkpoint :c
You should be able now, I have changed the type to SDXL 1.0
@Panchovix Nothing yet :'c
Civitai staff chooses the models available for the onsite generator
What are the differences in the perpendicular version vs the normal cyberfix beside the method of making it? Does it have better anatomy or do somethings better?
Why I can't use this lora
This isn't a LORA, it's a checkpoint.
@raeliz so how can I use it
You use it as your model. You need to load it into your front end of choice.
@raeliz but the problem is idk how to load it, use the prompt or?
@Onepersonabove Use ComfyUi, SwarmUI or ForgeUI to use Noob based models. save them on your checkpoint folder.
@norim it's all pc, I'm using mobile
Is it only available on PC?
@Onepersonabove you can use civitai's online generator. Alternatively you can try this: https://youtu.be/sWlb9tQMhQc?si=dxJL6vegz3MI2IVo
@raeliz ok tysm🙏
I am trying your model with Enigmata IONgrid lora. And so far I really like it.
But is there a way to find the artist list included.
So far i found Deadflow, incase, showasu_no_okina.
So i am trying at random.
Dataset list, try most artist with 100 to 300 pics if possible at least from Danboour:
@norim Thank you
CyberRealistic XL had released v5. Do you have a plan to update the merge model?
Could you please share your model merging workflow
很好的模型，要是更新到最新版就好了
What Sampler should I use?
euler a / normal = quick generation
dpmpp2s a / beta = best results
Every generation comes out with even lighting, I'm trying to make dynamic chiaoscurro style images but everything is flat lit.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.