This LoRA allows the generation of highly realistic selfies in typical smartphone style. It is optimized for natural poses, various facial expressions, different hair colors, lengths, skin tones, and body types. The generated images look like authentic amateur selfies, whether close-up shots, relaxed smiles, playful faces, or casual poses in bed, on the sofa, at the beach, or indoors. Perfect for realistic portraits, lifestyle, and everyday shots.
My-Settings Z-Image : Euler/Normal or Simple-Beta 832x1216 works great but higher resolutions are possible, for example 1280x1920
"Other resolutions or formats also work without any problems. You can try out what works best for you."
My-Settings Flux2klein9b Distilled : Res_2s/Beta 4-6 Steps 832x1216 works great but higher resolutions are possible, for example 1280x1920, 1152x1728
For Flux2klein9b set the strength of 1.25-1.50 (Distilled)
(I created my example photos with Res_2s/Beta 4 Steps and a strength of 1.50 to make the look look better on the distilled model. On the base model you have to test the settings yourself, I only use it for training the Loras.)
The 4B settings are similar, but you don't need to make the Lora's strength as high; 1.0-1.25 is sufficient. But for both models, use the settings that work best for you, these are all settings where I am personally satisfied with them :) just test it yourself.
Before anyone wonders why the 4b sample photos look sharper, the sample photos at 9b are in 1mp resolution and at 4b in 2mp :-D so please don't be surprised.
Description
For Flux2klein9b - base and distilled
FAQ
Comments (15)
Early thoughts on K9B vs ZIB/T?
Well, to be honest, I didn't want to switch to Z-Image anymore. I was a big fan of Zit, but klein9b is just incredibly good, especially because you still have the option to determine the pose from reference photos or insert clothes etc. and people, image editing, style transfer, there are countless options you have, that makes the model better in my eyes :). For me, Zit also has a big problem with anatomy. Klein9b, for example, does this really well. I'm already working on larger projects with nsfw and it works really amazingly well where Zit has gotten weaker and weaker. With Zit, the genitals always look so deformed and horrible that you have no desire to create a Lora. For portraits and not naked loras, I still think Zit is great and it learns damn quickly and well. Both models are great, but klein9b has even more to offer me, so this is my new main model. Best regards
And the speed is truly breathtaking. I use the distilled fp4 model and create 2mp photos in less than 2 seconds, which is really more than just intense and in such good quality.
I feel ya. Klein is much better than expected. But I just can't see to get what I want yet. Something is always a little off. I tried some Qwen prompts in Klein and they were awful! It's like going to less than 1/2 the parameters actually does make a difference. I'm hoping I just have a bad prompting style or could use a better workflow.
The only thing I tried to train was a weight slider and it did not go well. Works, but extremely softens the image. Like almost to anime detail. Hoping it's just me.
klein is the real goat imo especially the 9B edit one,
hope zomni or zedit can deliver same quality and speed atleast.
--
Anyway is there a plan for butt selfie for klein OP??
I will echo everyone here. I was SO excited when z-base came out. Turbo's character lora training was so simple and powerful that I figured base would be the same way. Of course I was wrong about that.
Klein has blown everything else away for me. The editing and T2I are both incredibly powerful.
@brnfd24434343d That's the amazing thing about Klein -- human-made prompts actually work insanely well. I hate using gargantuan LLM prompts, and ZIT seems to only work well with them. Seems to me the point of the CLIP should be to translate my prompt, not have another model in front.
@fronyax Yes, I am in the process of creating all my Loras from zit for klein9b. I actually wanted to upload 1-2 Loras today, but didn't get around to it. More will surely follow in the next few days :) Best regards
Does not work well with inpainting. Keeps changing face of the model to pretrained.
Yes, what do you expect? The Lora is trained for Text2image to create realistic smartphone-style selfies. I never mentioned that it can be used for inpaint or image editing? It's clear that Lora changes faces when she's trained on faces, that's self-explanatory.
Does it work with Flux2 Klein 4B fp8 ?
I quickly trained and uploaded a version for Klein4b. Have fun with it :-)
@bloodforce88 thank you so much. You are a true legend
@akringerzekrom656 No problem :) I hope everything works well
@bloodforce88 yes it's working really good. I tried in both Z Image Turbo and Flux2 klein 4B. I used gguf files in my GTX 1650Ti + 12gb RAM laptop and it is working very smoothly. This LoRa is very well optimised. Nice work