Krea 2 OSS - Optimized FP8 Weights (Turbo)

We need a fork of this one with fully NSFW capability XD

https://civitai.red/models/2725430/krea-2-nsfw-v3?modelVersionId=3071760 (Doesn't change the image you started with too much).

https://civitai.red/models/2728644/krea-2-mystic-xxx

It is uncensored. Watch this video, it shows how to get rid of restrictions. And the purest truth in that video - I tried it myself yesterday. I didn't try anything 'hardcore' because I'm not interested in it. But it produces the usual aesthetic 'nsfw' images perfectly.

https://www.youtube.com/watch?v=QC1BcvRyp1A

the idea is right, but the params are strange

it does work but the output seems drop a lot of quality and strange coloring

@xhamster Yes, I noticed that too, and I noticed that the texture of the skin feels more plastic.

i agree it works but the quality drops and character knowledge also gets lesser.

for me works perfect

How do you use it?

@chaank2013585 Connect this node to the positive prompt node and output node in ksampler.

Yes, maybe yes. The details are far away from Z-Image Capabilities.

I think yes. "From the box" Krea2 for me give better results than Z image of Klein. I made maybe 15-20 images with Krea2 turbo and all results was good. For NSFW... Krea accidentally created such an image. And as far as I could see, Krea has been trained on the anatomy of the upper part of the female body. I wondered - is it possible that the text encoder simply does not pass the request to the model. I am not an expert, but maybe after applying the corrected text encoder, the model itself will not even need to be trained. I don't know, these are just observations. But I really liked the model. I would say its photorealism is much higher quality than Zimage or Klein.

PS. This model handles text well. I tried English, Lithuanian and Russian texts. The model correctly represented both Lithuanian "žšąų" and Cyrillic. It even coped with bilingual text - Russian text on one line, Lithuanian on the other.

I'd say yes.. this is an 0.1 model, and it already rocks, we can just imagine what future versions will look like.

I'm a big Z fan but damn this is a good model and out the box its doing better things than what Z did at least for me. can't wait for things to come for it.

@MrGhost_27 @MrGhost_27 I spent a couple of hours yesterday with Krea2. The results were pretty much what I wanted. No 3 arms, no 6 fingers. It did a lot of things better than other models with special Lora's.

And if this is only version 0.1, what can we expect from version 1.0?

What about an IP-Adapter? z-image-turbo doesn't have an IP-Adapter or edit model right? It would be nice if Krea 2 did.

@herkus_baronas631 yeah the text is quite incredible, at first I thought that I was just getting incredibly lucky

Yes

yup, works well

Flawless even on a 6 GB RTX4050

Hi! Unfortunately, the base model doesn’t shine in terms of variety—but that’s expected, since it’s meant to be a foundation. We’ll have to wait for users to release their own fine-tuned versions, though that won’t happen right away. Usually, the first or hastily made community versions end up worse than the base model—everyone’s rushing to be first, and many don’t care much about quality.

In my opinion, it’s best to wait. This model has already proven itself superior to Z-Image in many aspects.

it is because of the turbo distillation, try the raw model and you see there is much more variety and dynamic lively poses. Turbo comes at a cost. https://huggingface.co/Comfy-Org/Krea-2/tree/main

and yeah i find Krea2 much better than ZImage, QwenImage 2512, Flux.2 Klein 9B.

I found this custom node. It work pretty well on a "Flux Preset" in Bold mode https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance

Not (yet), but support is planned.

It's supported now

Update ComfyUI. Open template, search "krea2", load workflow. If you get errors, copy the model names, Google them, download

https://huggingface.co/Comfy-Org/Krea-2/tree/main here

I guess the author tried to "aura farm"

@apo1s Ok thanks for clarifying but imagine you're the end user you have no idea who is who and which one you should get. So you must be the same person as AlperKTS but with a different name here and you're publishing the exact same thing here and on Hugging Face under different names. Do you think we just know that somehow, we don't.

What do you mean "soft skin", "soft images"? Give to us some examples.

@xNzX https://civitai.red/posts/29905351 Smooth skin, smooth weapon texture (no scratches or detailing on it). Tried a bunch of completely different image prompts, scenes, schedulers, etc.... This FP8 version always came out looking less detailed/soft compared to all the others (which is odd given it was coming out worse than the nvfp4 version consistently).

I have a large comparison of images when I tested (exact same seeds, etc), but I could only find exact same seed/prompt for nvfp4 & FP8 version right now (I'll need to sort them properly).

Yes, on your example obviously seen difference - less detailes on skin and weapon. But I use exactly your prompt and settings (workflow included) - and result is the same good, as your other (good) versions of Krea 2 - https://civitai.red/posts/29910666/ Here is result with simple (standard) VAE decoder. In VAE upscaled - even more details (must have).
So, something wrong with your file (??) Investigation needed. But I forgot where I download my file. May be from Huggingface? Or from here?
Here is SHA256 of my file: 2d3523507c59df965e5d4ec9a1b9b4591297a50c058915c36fb29d124d30f64e
Open cmd.exe from folder with file, then paste: CertUtil -hashfile krea2_turbo_fp8.safetensors SHA256

Here is exactly the same SHA256 of my file: https://huggingface.co/AlperKTS/Krea2_FP8/blob/main/krea2_turbo_fp8.safetensors

@xNzX I no longer have the file in question, because I deleted it. But I think I downloaded mine from https://huggingface.co/AlperKTS/Krea2_FP8/tree/main. I use SwarmUI, & when I go into the checkpoints tab there, I have to manually update the metadata by connecting to civitai servers to import it. It auto-connects/linked to this page when it updated the metadata info though; so I assume the checksum is the same one as on this page for it to have done that.

PS - whenever I want to go back to this page, or another checkpoint/LoRA; then I usually click the metadata link from the checkpoint file in SwarmUI to be sure it's the same thing (that's why I think the checksum would probably be the same).

Also, in my personal checkpoint list; I have https://huggingface.co/AlperKTS/Krea2_FP8/tree/main with a strike line through it. I download so many LLMs, etc, that I keep a personal record (& a strike line means I already tried it, & thought it was poor quality). There's no link for this checkpoint page here on civitai on my list either.

I was thorough in checking though; multiple different text encoders, with LoRAs/without any LoRAs, pretty much all the different settings (& starting a brand new session from scratch whilst manually entering the new settings, just to be absolutely sure). It was always lower quality across multiple different versions of SwarmUI though.

@apo1s The plot thickens. Only obvious difference I can see is that I'm using SwarmUI. I ended up doing another new install for it again right around the time Krea 2 launched, so it was clean with no modding by that point either.

PS - I changed my original comment to better reflect the issue is probably on my end.

@apo1s Please test EXACTLY model from link from THIS page: krea2TurboFP8_krea2TURBO.safetensors
And get SHA256 of that.
You test - 'ComfyUI Scaled FP8'. It's different file. User ShadowCell report "_scaled" is good:
krea2_turbo_int8_convrot <-- good

krea2_turbo_fp8_scaled <---- good

krea2_turbo_nvfp4 <---- good

krea2_turbo_fp8 <----- BAD

I also used these https://civitai.red/models/2775340/krea2-textfusion-refusal-reduction-lora?modelVersionId=3125118

https://civitai.red/models/2731465/qwen3-vl-4b-abliterated-comfyui-krea-2-text-encoder-bf16-fp8?modelVersionId=3070870

It's unlikely, but it's possible "beautiful" parsed as smooth, fashion shoot skin or something; but is also censored in the original krea 2 (it censors expressions, breast size, etc, & I noticed a lot of other silly things)....but I find it unlikely, as it didn't effect the other versions. A corruption during download is possible....but again very unlikely. The LoRA is only 26mb, so I'd maybe try an image gen with that & see what happens.

Seems the problem is likely on my end though. I'm tired as hell, & off to bed, but I can see about downloading it again tomorrow (if it's any help).

https://civitai.red/posts/29930049 https://civitai.red/images/137462988 There's another 2 images, taken maybe 2 weeks before the other ones, & with the LoRA removed (but just the the abliterated text encoder still in use....which I noticed didn't do much with the other versions, anyway).

Ok, so it appears the issue has been narrowed down now. I have another Forge Neo install (which I also installed right around Krea 2 launch) to test performance differences. It re-directs to my SwarmUI resources folder to use the exact same files. & I have images I dropped in there with the same prompt I created.... & the images using Krea 2 FP8 in Forge Neo look fine (freckles on face, scratches on gun, etc; as good as I'd expect them to be).

So it appears whatever caused it; was a bug in SwarmUI (which usually slightly trails behind updates for the Comyfi back-end).

https://civitai.com/images/137507138 (Forge Neo version)

Civitai doesn't auto-detect my resources for Forge, but I dropped the image into my Forge PNG info tab, & it shows them as same.

The seed is different (so let me see if I have two the exact same; but the rest of the images in Forge appear fine, too, anyway.....). I'm looking through all my Forge Output images; & none of them have this issue.

PS - if anyone else has SwarmUI, & could test the same image prompt to see if the same thing happens.... This should probably be reported as a bug.