Just prompt: switch to realistic style
Mar 07 2026: @astroaida reported that comfyui was complaining about some layers of this lora with this message:
ERROR lora diffusion_model.double_blocks.0.txt_attn.qkv.weight shape '[12288, 4096]' is invalid for input of size 16777216
...
This was fixed and a new file was uploaded just download the new file. The old file was removed. Description
FAQ
Comments (10)
In your examples, is the "no lora" the input image? or is there another input image, would be great if you could show the input image for a better comparison. Thanks.
Yes, that’s the input image. It was just my default text before, but now that you mentioned it, I realize it wasn’t a good choice of wording.
very well trained. Half good and half bad for transofrmation. Good is it really learned every concept, bad is it sticks too much to the original so anatomy is unrealistic with human transformation. Maybe lowering strenght can help. But for Fantasy, abstract etc. art style this is really an awesome lora
ERROR lora diffusion_model.double_blocks.7.txt_attn.qkv.weight shape '[12288, 4096]' is invalid for input of size 16777216 it gives this error both in fp8 and base models. it needs some fix i think. Dont know all the weights are working cuz pure klein can "make it realism too" Made some LLM chit chat and they say the lora is using different model architectur (like flux1) so maybe you uploaded wrong lora or you trained it with other model but tagged it Klein 9b by accident?
Thank you very much for reporting this.
It looks like ComfyUI changed something under the hood again. The LoRA itself is fine and was trained on the correct model, but ComfyUI has become a bit more strict about how some of the weights are shaped. I did quite a bit of tinkering with the architecture when training this one, so it ended up being a little different from what the loader now expects.
Nothing was wrong with the model though. I just adjusted the double block weights so they match what the current ComfyUI loader expects, and that should prevent the error from appearing now.
Thanks again for taking the time to report it and for the helpful comment.
@alcaitiff Love your loras thanks. for the update. I use your Unchained lora too. But the problem with your loras are you train them so perfect so we have to lower strenghts :D:D BTW can i give next lora idea for Klein?
😍 👍🏻
A plea for ZIT Version
How would that work? Flux klein has image editing capabilites, ZIT does not.
I've given this lora a good run the last couple of days and kind of having issues with faces mostly. I had the idea to try testing the lora versus having it off and tbh I've had better results without the lora.