Update:
- V3 is about half RAW still runs at 8-12 steps.
- Added V3 FP8.
- Added V3 Int4 ConvRot SR (stochastic rounding).
⚠️ IMPORTANT — Applies to v1.0, v1.1,v2, v3 ⚠️
❌ Do NOT use the INT8 Fast custom node.
This checkpoint is already in native ComfyUI INT8 (ConvRot) format.
✅ Use the standard ComfyUI Load Diffusion Model node.
Requirements
- ComfyUI v0.27+
- ComfyUI Kitchen
Krea2 Center v1: similar, experimenting with some block merging.
Krea2 Center v0.5 explores whether Krea2 can become more expressive through carefully balanced LoRA merging while preserving the model's realism, flexibility, and overall character. The goal isn't to replace Krea, but to gently extend its emotional range across portraits, fantasy, sci-fi, and stylized scenes.
Added Int8: about 25% faster on 50 series Nvidia don't have a 40 or 30 series to test.
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Description
FAQ
Comments (31)
V0.5 seems to work better. V1 get a lot more errors like nipples and nether region just melting through clothes. Not sure if that info is helpful, but my experience so far.
thanks for the feedback I will have a look did not have that issue except one lost “shirt”
I did test some prompts and yes its a bit much also some unsuspecting loras on top make it worse.
also if I use the test prompt I just made to check that v0.5 is about as bad as v1
@TiwazM could be just seed luck on my test prompts. Been doing lots of tests, didn't see as many on the tests I ran on 0.5 using same prompts. It's still the best checkpoint out there I have tried so far for krea2 though. Had lots of issues with other finetunes like moles and freckles and other blemishes going crazy, this model makes nice skin, both versions.
@i_suck_at_this_game lets hope I did not break it in 1.1 coming soon ;)
@TiwazM 1.1 so far gets a lot of wet results. Good for summer fun pictures lol.
@i_suck_at_this_game can’t say I have had that issue yet
@i_suck_at_this_game maybe I found what could be causing it if could maybe explain the issue
1.1 int8 convrot is also like this it seems. and the wet body lol
@TiwazM im gonna post a few of the gens that illustrate what we're talking about
@hughchungus420 I might have fixed it but yes with prompt ideally so I can check if I fixed it
@TiwazM Let me know if you've seen my example images, seems Civitai is still very broken and hasn't updated model pages with submitted images as of the past 24 hours.
@hughchungus420 I wish but as broken as galleries are at the moment.
feel free to message me some example prompts and I will test them
First of all, thank you so much for your amazing work. In my opinion, you are currently creating the absolute best Krea 2 finetunes out there! The quality is unmatched.
I noticed that your latest releases are exclusively in INT8 format. While I understand this provides a speed boost for NVIDIA users, it unfortunately makes the models incredibly slow and unstable on non-NVIDIA hardware (like AMD/ROCm or Mac).
Even though the non-NVIDIA community might be small, we love your work and feel a bit left out. Would it be possible for future releases to also provide a BF16 version, or at least an FP8 version? That would make your incredible models accessible to everyone again.
Thanks for considering, and keep up the fantastic work!
I will just checking on an issue probably for v1.2
@TiwazM Awesome, thank you so much for looking into it! An additional FP8 or BF16 version for v1.2 would be amazing for all of us on non-Nvidia hardware. Appreciate your effort!
Update: The latest version of GGUFY (v0.7.1) now fully supports converting ComfyUI Kitchen's INT8 ConvRot layout alongside all other common formats:
https://github.com/qskousen/ggufy
So if you are having trouble with the native INT8 ConvRot format, you can now easily convert it yourself into FP8, for example. Thanks to streaming and CPU support, GGUFY has very low hardware requirements.
But this is exactly why there is really no reason why checkpoints shouldn't be released directly in full BF16. This way, every user has the freedom to generate their desired target format in optimal quality on their own machine – including INT8 ConvRot.
@photophobia371 pretty cool tool I will have a look, as for int8 convrot I make them with my own node with comfy kitchen.
@photophobia371 as for bf16 I will think about it, would have done an fp8 already but got so sidetracked trying to make v2 and I dont release a quant unless I test it.
@TiwazM Thanks a lot for the super quick reply! On one hand, I’m really excited for V2 now, but on the other hand, I’m completely relaxed about it – because no matter which format you choose to release it in, I know I'll be able to use it.
Market demand dictates everything; with such a large user base, those who don't use native NVIDIA graphics cards for acceleration will have to upgrade their graphics cards.
I keep getting random noise as the output image, any idea why? all the other k2 models work with no issues for me
yes maybe are you using the latest Comfyui and the default loader?
@redlucario1735 I think I identified your issue check the model description...
@TiwazM You're absolutely right, that was the problem, its working as intended with the default node. Great work on the finetune btw.
@redlucario1735 Thanks glad it worked out
i think there's an issue with 1.1 INT8, 1.0 int8 worked fine with native int8 comfyui node, but with 1.1 i'm getting noise instead .With normal load node it looks mindblowing though, amazing checkpoint.