First attempts at creating a general finetune Lora for Klein 9B
Model now knows:
Pussy
Pussy spread
Anus
Ass Spread
Insertion
Gaping
Enhanced nipples knowledge
and way more
Also use amateur photo for a cellphone type of image.
Use the Lora weight of 0.8 to 1.2
This finetune doesn't fix Klein limbs issues.
Description
Initial Release
FAQ
Comments (21)
thank you,PLZ add Enhanced pubic hair knowledge
Does this Lora replace https://civitai.com/models/2322539/nude-legs-spreading-klein-9b?modelVersionId=2612784 ?
Yes. It was part of the training dataset.
Damn. How big was your dataset for this? Klein seems to take to it really well.
2000 images. Klein respond very well to training. Big potential there.
@razzz My only hesitation with Klein, is that it creates more body horror than other models of this generation, messing up things like hands when models haven't had much trouble with that in a while now. Do you think that can actually be significantly improved with training like this? I just assumed it would be unlikely. Very impressed with your examples, but just curious about your perspective on it.
@Jellai It can be improve greatly yes. I see better success rate with posed that it's been finetuned on. But It would take a bigger project to fix everything. I wanted to do that project with ZiB at first but I got bad results and it seems it's a known issue with big dataset. There's no ideal solutions, in a perfect world we would have ZiT speed, Qwen perfection and text and Klein styles, variety and trainability.
But for now I really enjoy Klein, it's pretty, fast and if you can deal with retries it's not that bad.
@razzz Yeah, that's where I'm at too. For realistic editing, you can't beat the ratio of speed to results that Klein has. Z-Image is surprisingly better at art styles, so I'm looking forward to Omni/Edit on that line, but it's hard to beat the total capability of Klein for realistic editing. You can make a few images in the time to do one test in other edit models, so even with errors, the iteration speed gets you to the end faster, and with really good results on a good dice roll.
A rare gem in an ocean of trash
I wish more people realized Z-image is trash compared to Klein.
i wish people would stop pretending there's only one use-case for AI and only one model is 'best' for everyone. like, is this two 12yos arguing about consoles, or does anyone figure "hey, maybe it's my approach and intent toward a model that matters more, and not just whether the model suits exactly what i do."
it depends on your use case and what you want to achieve... there is no "best" model.
Any reason why cfg is at 1.1 (and not 1.0) in the second KSampler in your workflow??
Also FluxGuidance node on/off does not make any difference in the final result..
cfg 1.1 will boost saturation a bit.
This seems to be a bunch of loras merged together. That's still a lora, right? You are calling it a fine tune, but isn't that a whole model all inclusive?
This is one big training with 4000+ images and multiple concepts done on a 6000 pro. Lora is the Low-Rank technology to make it more efficient but this is still a finetune.
@razzz How is this different from https://civitai.com/models/2374977/realism-engine-klein ?
@razzz , what would you say is the use case of this one vs your Realism Engine LoRA?
It's woman only and less concepts too.
@razzz So, Realism Engine's Dataset includes this one's?
@pretty_pixels Pretty much yeah