v1.0 - changes anatomy and overall picture quality
v1.5 - only changes anatomy
4B - only changes anatomy
This concept slider LoRA was trained to correct the anatomy / quality issues with Klein models.
The LoRA can't fix everything but it helps. It was basically trained so that negative weights have the effect of a negative stable-diffusion prompt, and positive have the opposite effects.
Weight 2.0 mostly keeps the seed intact and can be used to fix minor glitches, 3.0 fixes some more prominent artifacts.
The LoRA also improves a bit the overall quality of the picture, makes the picture more "clean".
Note on the 4B version:
I never managed to generate good quality images without the LoRA, so I didn't post images for it. I was able to confirm that the LoRA behaves similarily to the 9B version, but I probably won't update it unless the bad quality of the 4B model is due to my workflow.
Hope you like it!
Description
4B version, I never managed to have a good quality image with 4B so this is untested.
FAQ
Comments (45)
your examples show better anatomy sure but the images gain a very ai-ish fake sheen . the extra DOF doesnt help realism. feels like a step back to early flux gens
Ah yes this isn't a realism LoRA, I don't really care about an AI image being AI-ish, but I prefer when the generated persons have only two legs.
There, fixed it in 1.5.
@l226 wow you rolled that out fast. yeah that looks a lot better. doesnt ruin the realism anymore
@QualityControl It's Klein's game to lose now... Some of the loras train in an hour or two. Look at the file size, 20MB.
@l226 What step count were you testing this at?
@ZootAllures9111 the sample are generated in 4 steps, if you mean the training steps it's 50 for 1.5, 25 for 1.0
@l226 OMG NOOO
could you do a klein-4b version please
Done, I'm not sure how to use Klein 4B for T2I though, each time I generate a person the quality is awful... So the 4B version is not tested, I hope it's good.
@l226 The 4b is used the exact same way as 9b but with the qwen3-4b text encoder instead of the qwen3-8b used with the klein-9b model ,
I'm using the fp8 version of the model and the quality Is decent enough but the anatomy is not with extra or missing fingers and limbs , that's why I thought is type of lora may help.
thanks for the lora, will test it and see if it improve things
At least it is not the level of SD3
@qek that's true , for me it's a huge step forwards since it's works good enough as an edit model that's not too big and is fast enough for lower end hardware , and since the model is apache-2 licensed I expect to see in the future full fine-tunes of the models that may fix it's shortcomings
@qek Ok ok so I do have an issue with my workflow then, the quality is way worse than SD3 for me...
I tried version 1.5 out on Klein 9b distill. It does seem to help address anatomy issues. I have the strength up quite high -> 3.0. And I'm using it with other loras but based on some initial testing I'm pleased. Thank you for sharing.
Same here. Flux 9B Klein. Does not help no matter what setting 1.0 - 3.0
why so fancy with unintuitive iteration nomenclature? hurts my brain a bit
Doesn't take much then ;)
Here you go, I hope your brain stays safe for a while!
@l226 Huh, what now? Uh thanks for the fix and why should it take more to hurt a bit? Did i stumble into a hidden contest of sorts, some scientific question to be answered, "how much does it take to..."? Confusing it seems....🧐 Anyway, thanks for the "low effort" i guess? 🙃
@eyeonyou it takes little to hurt you
ziamge base plz
Thanks man! And 4b can produce just need to run it with loras. but this anatomy lora is working well!
Do people really need this? I almost never get morph monsters.
I don't even need negative prompts to begin with. You guys must do something wrong...
Some prompts are fine, others frequently anatomical horrorshow. Depends on how many positional instructions i put in. Maybe? Dunno.
I was skeptical in the beginning, but after testing for a while, I can confirm this lora actually works for editing images. And as another user mentioned, you can use a strength of 3.0 to remain on the safe side, or you can explore.
Really helpful and saves a lot of time, when you just can use the lora instead of doing inpainting/repair work. the overall image stays the same and the anatomy error gets fixed. Thanks a lot!
Is it possible to have a version without altering the original face ? I think it is getting changed everytime in v1.5
It's hard to have a LoRA that "fixes body parts" without having side effects on other body parts unfortunately. I think these are minor changes, and I don't want to spend a ton of hours perfecting it, sorry!
@l226 remove any faces in your dataset I think
@l226 It's very major
@chasingexp245 Ok then don't use it...
@Sassyphrassy I don't have faces in my dataset :)
works really well indeed
请问能否添加更多的训练数据集来增加lora质量，对我们很有帮助
Thank fk, I was going crazy with Klein models, is there an issue with their anatomy? All my dudes have penises coming out from their behinds.
I had that problem recently when I was fighting with Z-Image Turbo. It was obsessed with having at least one hand grasping the penis. Usually the penis would be disembodied with someone holding it in their hand. I don't remember which loras I was using but it was very frustrating. So, I tried prompting that the person had their hands tied behind their back. So, Z-Image Turbo responded by helpfully having the penis protrude from the butt or just above it, with, you guessed it — a hand grasping it. I finally gave up and switched back to Wan 2.2.
I'm seeing this again and it still cracks me up. The first response I had written out was probably a much better explanation. It was definitely funnier. Question about your dudes: Is there a moon with a dude with a big rocket crashing into a crater from behind? Or is the dude crashing HIS big rocket into a moon's crater? In the first one- that's not the dude's rocket , it's just "a big rocket", and the model is like "I need to stick this big rocket into a crater, and that dude's just standing by that moon doin' nothin. Let me just stick this in there big fella and... PERFECT!", Now you've got a dude with big rocket stuck in his butt and some pretty confused astronauts in the wrong moon crater. Basically: Give ownership of the rocket to your dude, make sure they're headed to the correct planet, and I guess "directional coordinates" like "upside down in the top right at 45* angle"
This makes an incredible difference. Thank you.
How did you set it up to make it work for you? I enable this lor 0.5 - 3.0 and nothing changes at all. Never fixed anything for me. What is wrong ?
do your sliders work on 9b base by chance?
Klein4B "Legs are good, so more legs is better"
Does Laura have issues with face changing? Many Laura systems alter faces, which causes the entire workflow to fail. If there is this issue, please point it out in the description.
Who is Laura?
Klein 9B v1.5.
How to make it work? It never works for me.
It has zero effect and 3 fingers or 6 fingers still generated, as well as 3rd legs.
Same here, it does not makes any difference, anatomy is still effed up.