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
After training with an additional 200+ images I think this is very close to being done so I'm labeling it a release candidate. I'll still try improving it further and incorporate any feedback I get.
Sampler/Scheduler: euler/beta, dpmpp_2m/simple, er_sde/beta57, sa_solver/simple
CFG: 1.0
Steps: 5-10
The workflow is embedded in all of my images.
P.S. I accidentally incorporated a model that doesn't allow merges into this version, https://civitai.red/models/2786809/krealism. However, the creator of that model, @lonecatone23, has graciously allowed me to leave it in this version. Please check out Krealism. It has kind of a softer SDXL-like effect that has been very nicely refined. I used it in this release to moderate the over-sharpness of the Training Base and then layered my private loras on top of that before remerging again with my excellent-yet-flawed v1.0.
FAQ
Comments (50)
As was aptly said to you by a more skilled commenter than me some time ago, GBRX you magnificent bastard, you’ve done it again. This RC candidate is a masterpiece!
I agree this guy know his stuff
Thank you so much for the update! Great job done, really one of the best models (in my humble opinion). My favorite. NSFW content is much better now. Renders without any graininess or blur, with some LoRAs it rendered poorly, but not with your model. Thanks again!
Awesome, thanks!
Thank you sincerely for releasing this new version. I have used most of your models and have been very satisfied with them. However, there is one slight drawback with this version: the skin tends to look a bit "plastic" when zoomed in, compared to v1. I feel that v1 achieved a more realistic look, so that is a bit of a shame.
I have to admit that v1.0 is my favorite, even though it has flaws. With these new versions I'm trying to remove the v1.0 flaws while keeping all the benefits, if that's even possible.
Thanks for the effort you put into your checkpoints and for not hiding them behind a paywall... I'm going to try out Krea right away.😁🤜🤛
Any chance for an INT8 ConvRot version?
Yes, there is one. Check the download options. There's BF16, FP8 and INT8 convrot.
Many thanks for the perfect models.
However, the v2 training base version is more accurate than the v2rc1.
Thanks for the feedback. Do you mean that v2 TB is better at prompt adherence than v2 RC1?
@GBRX Sorry for my English—that’s exactly what I meant. For example, age specifications for characters are depicted better or more precisely (based on the prompt, of course) without LoRAs. But these are really just fine-grained observations.
Please don't get me wrong.
@Woopy123456 Your English is perfect and I greatly appreciate the feedback. It helps me make the model better in the future.
Thanks for your work! I bought your WAN workflow recently — would the KREA 2 model be a good fit for it? Any quick tips or recommendations would be great, thanks!
Hey, thanks for your purchase! Krea 2 makes excellent reference images for videos. It's probably the best model right now for that purpose. Just remember that the reference image is going to be the first frame of the video and it's probably more important than the video's prompt. WAN will infer a lot from the content of the reference image so it helps if any 'action' is already happening in the image. Then you don't need to work as hard on the prompt.
@GBRX for generating good NSFW photos of the first frame, would your checkpoint and the Krea 2 workflow — the one inside the photo examples in your profile — work for me? Do I need to connect any LoRAs, or is everything for NSFW already baked into the checkpoint?
@proggmannme423 That's right. You just need the model and the workflow. No loras are needed, although it doesn't hurt to check out some of the Krea loras to see if they help. Otherwise you should be able to accomplish most content styles with just prompting. If you're strictly doing NSFW then you may want to use my original V1.0 model because I think that's what it excels at.
i dunno why but every time i prompt "hairy" and I specify the girls arms and legs typically get pretty hairy although I wasn't prompting for them. this has been since chroma too so not sure what im doing wrong
I've noticed that too sometimes. I think it must be inherent to Krea because I definitely don't have images like that in my training datsets.
@GBRX Thanks for your reply and the model. Been checking them out since Chroma. I'm starting to see it more with the FP8. The BF16 hides it behind more detail on the image. I know you've said the RCs are checkpoints before the full model so I'll keep testing.
@staticOne Ah ok, thanks for the feedback. Unfortunately, if it's just something that Krea tends to do then it will probably take a lot more training to get rid of it, but I'll do my best.
Man, this blend with mine is insane. I may like it more than Krealism by itself 🤔
ATM the best Krea 2 Checkpoint for Vehicles (which is always a Priority for me)🥰
thanks for the Buzz😊
Yeah, I used to use Flux for creating conceptual vehicles and it was pretty good at it. I need to try that with Krea.
@GBRX first it was ZIT, relatively decent vehicles; then came Flux 2 Klein, which was even better, but not even Seedream can keep up with Krea 2 - plus, so far I’ve only found three cars that Krea 2 didn’t recognize, and I have a pretty long list of cars from my 3D days😉
It's a very good overall checkpoint but I noticed that v2rc1 seems to have lost some of the Krea 2 knowledge, i.e. characters / people with specific ethnicities. It also has a tendency to create a semi-close up shot with the face as the focus and the body foreshortened, from a higher angle, which sometimes is hard to overcome with the prompt details. It may be related to my prompts specifically but the same thing does not happen to me using Krea 2 turbo itself.
The good things - the details, sharpness and realism are exceptional, and the female body representation is really great and fits my taste. The colors are great but sometimes slightly burned/too cinematic.
Overall incredibly good checkpoint.
Thanks for the thoughtful feedback. I've noticed some of those issues too.
a chance for ggufs? fp8 and int8 takes ages to render for me with 16gb vram. or have i misses a comfyui startoption? thx
I should be publishing the final v2.0 soon and I'll include GGUF versions as well. 16GB vram should be enough but you may want to set the clip device to be cpu in the Load Clip node. It might help or could be worse. Unfortunately Krea is just really slow unless you have a 5090 or better, which I don't.
Try also to use 4-bit TEs with fp8/int8.
Here you can download nvfp4 version:
https://huggingface.co/DreamFast/Qwen3-VL-4b-Heretic-ComfyUI/tree/main
int4 ConvRot version of qwen3vl:
https://huggingface.co/Merserk/qwen3vl-4b-int4-convrot/tree/main
https://huggingface.co/Skydexter/Krea2-Turbo-int4-ConvRot/tree/main
Also you can try a hybrid int4-int8 ConvRot version of Gonzalomo Krea v.1.0 by Lockout from HF:
https://huggingface.co/Lockout/krea2-comfy-int4-mixed/tree/main
@mphobbit ty, i will give it a try :)
My friend, you're absolutely right to keep working on Krea 2. I was truly sad to stop using your SDXL models, because they were the best for NSFW, but since this Krea 2 model came into my life, I can't use any others, because it gives you everything. Thnks for your work.
Yes, Krea is the first model I've seen that beats SDXL for NSFW in most circumstances
couldn't figure out good settings and tile sizes for upscaling with Ultimate SD Upscale yet. could you recommend an upscaling method that adds details?
Personally I don't do upscaling very much. With Krea I've found that just generating at the final resolution you want works best.
yeah that's what I am doing as well atm. longest site 1920 images are very detailed already. but it also affects the physique, making the character more slender. which is fine as it is possible to generated in lower res and then hiresfix into 2k to get different body physiques
@YaraAI Thanks for the tips. I'll play around with upscaling a bit to see what I think.
Krea 2 versions are great! Just one question, do you rec turning off other LoRAs like realistic snapshot or the detailer lora?
I'm not familiar with those krea loras so I can't say but it's certainly worth trying them and see what happens.
Love it! 1.0ALT plays nicer with Character LoRA (compared to the 2.0 early builds)
Hi there! I'm new to the world of AI image generation... I'm having trouble using checkpoints in ComfyUI. Could you help me figure out how to use them?
@arthuraraujok Use Gemini or Claude - they can get you a base workflow (or find one here). I am new to Comfy too
Hi there! I'm new to the world of AI image generation... I'm having trouble using checkpoints in ComfyUI. Could you help me figure out how to use them?
Just take the workflow from any of my images. That will show you how it's set up. Just open the full image and then drag it into ComfyUI.
Very good images but it dont play well with Loras
Which version(s)?
@GBRX V20RC1. I tested it. Default Krea 2 is sharp at strength .8, but very grainy on V20RC1 and slightly grainy on V10ALT. Character lora was trained on Civitai, 2000 steps.
@Hamster4Doom Thanks, I'll be releasing the final v2.0 soon and hopefully it will do better. It's not actually based on RC1 or ALT.
This model is great, pretty good NSFW prompt adherence and clean detailed outputs. Is there any chance for you to share a Krea2-raw (non-distilled) version so we can apply the turbo lora of our choice or even directly train on it? Alternatively, sharing a lora version of this that could be used on Krea2-raw would also be fine. Thanks!
Thanks for the great feedback. I haven't messed around with Krea 2 Raw yet but I do have the full version downloaded and plan to do something with it soon. As far as a lora version for this model, I'll take a look at that as well. I'll just need to make sure that it's good quality and outputs the same images at full strength.