Thanks to all of you making GonzaLomo Krea 2 such a popular model, it is now the flagship model used by generateporn.ai! If you are willing and able to support me some more then I would very much appreciate it if you purchased a monthly or yearly subscription - and paying with crypto would be a bonus as well. I don't charge for my models so even subscribing for one month at the lowest level would help a great deal.
Note: There are two FP8 download options for many versions. The 17GB option is an all-in-one checkpoint that includes the text encoder and vae as well, so load it like a regular checkpoint. The 12GB option is just the unet so load it like any diffusion model.
So many new model architectures, I can't keep up. Krea 2 is definitely going to have a place in my toolbox though. It might be the best current T2I model.
For best results, follow the recommended suggestions in the About This Version section.
v1.0 seems to be very good at NSFW content but it has the problem of producing unprompted nudity.
v1.0 Alt has addressed the unprompted nudity issue to some extent but is weaker at NSFW.
v2.0 Training Base is my test bed for new training.
v2.0 RC1 is a potential release candidate that builds on Training Base.
v2.0 Final is based on the original v1.0, abandoning the ALT, Training Base and RC lineages completely.
v3.0 is just a refinement of v2.0.
Beginning with v2.0, there are at least 5 versions available for download:
BF16
FP8
INT8 convrot
Q4_0 GGUF
FP8 All-in-One Checkpoint (text encoder and vae included)
Description
This is a new one for me but I thought I'd publish the model that I'm using as a new Krea2 training base in case anyone wants to play around with it. Basically, this is something I created that works pretty well on its own but its primary purpose is as a base model that I can apply a lot more future training to.
So, once trained, I'll release a final v2.0 that can fully replace the flawed v1.0 versions as a general purpose model.
Although, if you just care about NSFW content then the original v1.0 model is pretty awesome I think.
Sampler: euler, dpmpp_2m,, er_sde, sa_solver
Scheduler: beta, simple
CFG: 1.0
Steps: 5-7
The workflow is embedded in all of my images.
FAQ
Comments (28)
The only two type of models I am using at the moment are anima and Krea2. Your V1 is the first Krea2 I have ever tried. My pc struggles with it but somehow just manages 😂 and the wait is worth the result (it loads and unloads the model every time on comfy 🙄 need more than 12gb vram, I’ll fix it somehow).
I have spent a fair amount of hours messing about. The model seem very sensitive to prompt (love it).
The skin texture is just remarkable, including other parts of the body with “skin defects” making it even more surprising!
Finally something that doesn’t have the usual flux look.
Like you mentioned it likes to expose even when not prompted to do so, occasionally I have noticed that adding more clothing details helps with sfw. In my opinion we “need to learn” how to “speak” to the model. I have managed to have sfw, just need to be patient.
I haven’t tried the other two yet, but I will drop something else if I find anything.
Overall, every single time this model surprises me. Every new image is just wow!
Hope you will have better luck (for what YOU want to achieve) with V2.
If you want a nsfw model concentrate on that, if you want a sfw model do that. If you do both I think you will end up with a compromise and I think people will have something to say… again 😂🙄
Sorry for the long comment, just wanted to thank you and letting you know, this model is so good, I wait over 5minutes for each generation and despite that I am trying to experiment with it in the hope to help within my little.
Do a test, let me know how you get on with over-promting dressing a person and see if this would fix the v1 issue, again, it may be like I mentioned us needing to prompt correctly. 👍 (hope this helps)
Thanks so much for all the feedback. I also like the original v1.0 best and usually just try to deal with the nipple issues. I would like to create a version that isn't a compromise, that keeps everything I like about v1.0 but removes it's fatal flaw.
@GBRX I am dealing with the nipples too 😂😂 don’t care it looks too good 😂😂😂 one thing I am noticing is that it likes hairs (down there) is it just me being special or has been set up like this?
@123AccIdeTE I'm not sure if the hair is due to my training or whether it's just inherent to krea. I don't think my training dataset is really the cause but you never know.
Love the model. Tried the int8 version of 1.0 and comfy spewed warnings about unexpected layers or something. I quantized it myself from the bf16 using the comfy-model-tools script and it no longer did that. Just a heads up. Testing with a few gens using the same seeds they do indeed render different by a noticeable amount.
Thanks, I realize there are some issues with my int8 conversion. I'm working on fixing that.
@GBRX Thank you for the work you have done for the community. I am not sure which script you are using, but if you are converting from bf16 to int8 using ctq then I believe you can resolve the unet error issues by using the conversion command ctq -i KREA2MODELNAME.safetensors -o KREA2MODELNAME_int8_convrot.safetensors --int8 --scaling_mode row --convrot --convrot-group-size 256 --krea2 --simple . Using --comfy_quant , --save-quant-metadata , and/or --flux2 can cause the unet errors your int8 models are showing. I have personally tested two conversions and I was experiencing the same unet error issues because I was using the wrong command. When I dropped those 2 flags and switched from --flux2 to --krea2 the unet errors disappeared.
@rpgnut2910 Thanks for all the info! I'm using the INT8-Toolkit nodes now in ComfyUI and I think that it's working properly now but please let me know if you see any issues with v2.0 RC1 or later.
Seems to be good. No more comfy complaints at least.
I really, really want to love this model, but prompt adherence suffers on all versions, unfortunately. Tags-like prompting doesn't seem to help Krea2, unfortunately.
I think natural language prompting works best with krea but I think krea is not as good at z-image with prompt adherence. However, I'm slowly learning how krea likes to be prompted and it does seem to have its own quirks.
@GBRX I'm sure the text encoder being a censored LLM contributes to a lot of issues with adherence and training too. Heretic versions, unfortunately, don't always help.
@dobomex761604 Yeah, I find the results from heretic text encoders to be really hit and miss too.
Great work as always GBRX! Any hints on training with the base? Do you use aitoolkit or onetrainer? I have a great dataset but when I tried converting your bf16 base to diffusers directory to use with OneTrainer and training it, the Lora had grid pattern at all strengths and smoothed out all skin details at higher strengths. Thanks for any advice!
I don't normally use any of my models to train loras. I usually use the official model to train any loras. I think that gives the lora the best chance to be used with any model.
@GBRX cool, that makes sense. My Lora’s do work great with your models, will stick with training with Krea raw.
Wow, my friend! Longtime fan here! This Krea2 training base rocks HARD!! Super fun, great prompt guidance and a few surprises here and there! Well done! Keep on rocking! :-D
Thanks!
Beautiful model You could settle for me with an nf4 model to be able to run it on my rtx 5060 8 vram. Thank you
I'm still learning some of the new formats but I'll see what I can do.
I really feel Krea2 is best public image model around (finally a replacement for SDXL right?) but I'm a bit 'scared' (for lack of a better word) about the split chin/cleft chin comeback in many quality merge or trained models such as yours (overall I still like your model) and by comeback I mean compared to the base raw model at least.
I haven't noticed any obvious instances of flux chin. Can't you point me to an example?
Nope...this is not a replacement of SDXL... It can't work on "potatoes"
@konsai Yeah, it does replace almost all the functionality of SDXL but it is a lot slower and uses tons of vram.
@GBRX Yeah not as present as in Flux for sure, but for example <https://civitai.red/images/136999008 (and there are others, just gave the first one from one I've seen of one of your example images)
@VariousDog9932615 Ok, I think if that was the default or dominant chin type then it would be a problem but I'm not noticing that. It's a valid chin type that occurs in the real world so it only makes sense that it shows up to some degree. That being said, I'll try to make sure my lora datasets don't have too many of those chins.
Kreia2 is the best today)
I agree!