@iamddtla 出了这个lora了吗

@iamddtla 好吧，谢谢您

@iamddtla 我测试出来了，当提示词包含：对眉毛，眼睫毛，冷色系光照，滤镜，某某风格妆容等等之类的描述时，就会变成2.5D风格，emmm...感觉有点限制了发挥。

@iamddtla 好的，非常感谢您的帮助。

@iamddtla I would be happy to try to do the conversion. There is a tool called ggufy that can make int8 convrot files for krea. I don't think it requires that you can "run" the bf16. I have done a couple conversions with it... https://github.com/qskousen/ggufy

我没有特意描述眼睛，所以一般它是闭着的，如果我提示眼睛睁开，出来的图是睁眼的

generally Krea2 seems to do only one type of pussy. Can't do innie pussy and they all look kinda the same. Also it has problems with really big tits.
But otherwise the model is really good.

@iamddtla Thank you for letting me know. I understand the limitation. I really appreciate your work, and I’ll continue using the fp8 version.

@iamddtla If you want, i can convert a INT8+Convrot version for you to add here

@iamddtla sure that's okay, if you need in some moment just DM me, ComfyUI already integrated it natively, i just tested int8 version of this checkpoint and works the same as fp8 but generates in the half of time with no loss, and thanks for your contributions to this community :) legend

@simonishere correct me if im wrong, but to create the int8-convrot model, you need to run the conversion script on the bf16 version.

@DigiDr I'm not sure, i simply used the INT8 nodes with save model node and converted this model to INT8, then i tested and it works, 100% faster and same quality, tested on 3090

@simonishere any download links pls？the int8 is way too faster than fp8.

@simonishere Which int8 nodes, Bobs Flux nodes? TIA

@simonishere How did you do that? I can only get a model that generates noises. I used ComfyUI-INT8-Fast nodes to make int8-convrot version on my 4070s.

@DigiDr I think you're right, we can convert model to int8 only if we have a fp16/bf16 version.

@iamddtla A lot of us would be happy to to the conversion for you however to do it best would require us to have the 16bit to start with. This way the final output has some of the better quality weights to keep. Yes there are ways to make the fp8 faster but that approach misses out on the better quality you can get starting with the 16bit. Please take a look at https://github.com/qskousen/ggufy

@iamddtla FYI. Int 8 convrot is supported natively by comfyui now. (just have to update)

@DigiDr  I think it can be done on the 8 bit, however you leave a lot of quality out. And there is a difference. https://www.reddit.com/r/StableDiffusion/comments/1ut2rvc/krea2_bf16_vs_fp8_vs_int8_vs_gguf_vs_mxfp8_vs/

@iamddtla I recall you mentioning the 12GB card.. If you have some significant system ram in most trainers if not all of them you should be able to offload to system ram so you might be able to make it work (just takes longer for the offloading). Could I also suggest training on something like runpod? Or another approach would be as a lora so people could then use whatever base model they want.

@iamddtla I also recommend you lightning.ai where you have free credits every month to use high VRAM cards like L4 24GB for about 44 hours.

@xmsthc think you meant https://lightning.ai/ will need to look at them myself... never heard of em.

@heartsinaustin Oh...yes it is, I corrected it.

@xmsthc Just fyi to get the free credits at lightning.ai you will need to add a credit card.

@heartsinaustin No I didn't, I created my account without adding any card last year and still get free credits every month now, maybe they changed their TOS...

No I checked their doc, still card-free to use: https://lightning.ai/docs/platform/build/ai-studio

@xmsthc Your experience is different. after I signed up it did not give me free credits. I didnt screenshot it but somewhere it said add card to get free credits. So I did. However I haven't used any of the free credits yet. Want to confirm they exist before using something and getting charged more than I bargained for.