After nearly four months of training, testing, and refining, our first major Krea2 model is finally finished—and available completely free here on Civitai!
BabeGen WRC8 was trained primarily on H200 GPUs and is especially strong at:
Highly realistic, photorealistic images
Fantasy characters and environments
Natural skin, lighting, and fine details
Cinematic compositions
Creative and expressive prompts
Recommended settings
Sampler: ER SDE
Scheduler: Beta
Steps: 8
CFG: 1
Those settings are a great starting point, but feel free to experiment and share what works for you.
This has been a huge project for us, and we’re incredibly excited to finally release it to the community. Download it, make something awesome, and please share your generations—we would genuinely love to see what you create. ❤️
Have fun! <3
EDIT: You can NOW get prompts from any image you see on discord using our plugin https://github.com/pixelgraple/KREA2-Vision-Suite and also edit any prompt you see on discord using our qwen 3.8 model API <3
Don't want a discord plugin? use https://krea-vision.com !!
https://discord.gg/79pwmpVsyQ
see you there!
Description
FAQ
Comments (37)
BF16 version please
Sorry we cannot do that
@kaylabytes Why, if I may ask?
@CupRunethOverWithLSD @CupRunethOverWithLSD Original FP16 → quantize → INT8 → dequantize → reconstructed FP16 - the quality will always be the same as the original int8
Are you saying that you've trained already quantized version and you just don't have your BF16 model?
Converting from INT8 to BF16 won't make it pure BF16.
@kaylabytes That’s correct. So give us the full model please, we can make it by ourselves…
@Unicom the 16 version won't be released to the public nope
@CupRunethOverWithLSD the 16 version wont be released to the public sorry no
@kaylabytes I have no more questions for you and your model.
@kaylabytes Ok, too bad. Quants are useless for me.
this looks incredible, thank you
thanks!
@kaylabytes It is a very good model, you should be proud of your work
@Complement_Cascade thank you! <3
Thank you for your hard work.
This is a solid model! Really well-balanced with no strong biases. It follows prompts amazingly well, the colors and sharpness look great, and character LoRA fit seamlessly. You can tell a ton of hard work went into this. Thanks for making and sharing it!
thanks!!
This model is crazy, knows lots of concepts and follows the prompt perfectly. Thank you, kaylabytes. ♥
you're wecome!<3
Love the style and prompt adherence but sadly this model messes up hands as bad as early SDXL models somehow. I have to get really freaking lucky to either have hands outside the picture or even more lucky to get hands that have the correct amount of fingers or aren't just a mass of random flesh and bones.
i have never had hand or finger issues, at all with this model.
I mean, dude, you can look at the samples and see what you're saying isn't true. Are you using some controlnet to "fix" hands by accident? Or some obscure sampler?
@DaddyWolfgang no im not
Can you release SVDQ? Int8 is too resource-intensive and slow
INT8 is what we released, its not slow and not really resource intensive, runs about 14GB at peak for us.
@kaylabytes the SVDQ is twice as fast and the checkpoint takes up about 7-8 gigs. I already read that you don't plan to upload BF16, but then I need to vibcode scripts to convert from INT8 to FP8 myself, because INT8 can't be straight quantized to SVDQ. Could you at least upload FP8?
@deGENERATIVE_SQUAD i sure cant, we never made anything other than INT
@kaylabytes got it, it's time for a vibecode then
@deGENERATIVE_SQUAD the quality you will get from doing this will be crazy low - you realize you "can" go to fp8 from int8 but it will suffer extreme degradation - then you are also going a whole new arch..
@kaylabytes everything is under control, dequantize does not add new information, but I still do not need new information when quantizing in SVDQ W4A4, which is equivalent to int4 which preserves higher precision.
Love to try it, but 8INT only works for nvidia cards. Regular 8bit would be appreciated.
can always try it on babegenerator.ink