Contemporary Korean beauty with a luxury editorial edge.
Help me creating more models , add a few ractions on the posts, Post something down below even if You want to delete it afterwards, as a little push for the buzz hungry trainging system
Description
FAQ
Comments (11)
any chance you're planning on doing a Flux.1D version of this?
oh, I think flux is totally obsolete now and I have deleted my entire Flux Gym set-up , Krea2 is better at everything while runs faster and easier, every flux models I will upload would be the ones that I have already trained and stored.. which doesn't include this very model, I am sorry.
no I totally get it. I haven't tried Krea myself - maybe that's what I'm supposed to be looking at next. Thanks for responding
@tripjacks definitely try to migrate very soon , You can run it even on 6GB Vram and with only 8 steps and it still results in better images than Flux : it doesn't mutilate fingers, feet and anatomy as Flux does half of the time, even creates very challenging dance, gymnastic and acrobatic poses, natively understands and contains a lot of artwork styles , It's base model is also uncensored. Even Flux 2 with 60GB nonsense isn't half capable as Krea 2.
@NeuralVandal Appreciate it! I actually just got around to trying out your setup for Krea 2. It's weird - I used the exact settings you implemented in the first image. It looks similar but is of significantly lower quality - not sure why but I guess I'll keep playing around with it. Is there any chance you have plans on releasing the Jeonyul lora you have for Krea 2? It's quite good in flux. either way - appreciate the quality work you put out
@tripjacks hmm Are you using the same sampler and base model ? Also remember to update comfy ui and requirements like pip , I know a lot of people prefer to stay on 2.6 but I personally getting better results with the latest release . If problem still is persistence try different sampler (Eualer+Normal) and switching profiles on Krea2 enhancer node or bypassing it entirely. Plus : If you're using my workflow you need to install 2 nodes manually, I don't know if they're implemented in the comfy manager yet but I installed them manually , the guide is on my "Krea 2 compact T2I" workflow model page.
@tripjacks Yes I like her , I'd eventually do that.
@NeuralVandal Yeah super weird. I had to reinstall comfyui because the older version I was on wouldn't allow Krea 2 to run and it also wouldn't let me update for some reason. I used the exact same settings including the seed as a test run - had to install some nodes that were there but yeah. Super weird because it should have generated the same exact image. Also tried increasing steps from 8 to 25 but still no good overall. I'll also refer to your Krea 2 workflow guide and keep trying things. Appreciate the responses! I'm sure ill figure it out if I keep trying stuff I just haven't had the time
@tripjacks It might have been the user cache inside your venv folder, also to my knowledge : increasing steps over 12 on turbo model has diminishing returns
@NeuralVandal Well I figured out why. It's because I was using the wrong model lol. When I followed the model linked, initially it defaulted to the latest version which iskrea2_raw_int8_convrot. Took me too long to see that you were using krea2_turbo_fp8 instead. Once I switched to that everything was fixed. You weren't kidding about it way outperforming Flux. I'm shocked at how short the generation time is and the quality of the images. On Flux I'd do batches of 6 or 7 at 25 steps where a lot of them would be throwaways because of body mutilation and such. Nothing like it here. Really appreciate your responses and trying to help me with everything
@tripjacks glad it worked, have fun!