v1.3 - this is not upgrade, it is bigger and weaker, but maybe works better with other LoRA, if you want stronger better to use older versions.
🚀 V1.2 - Release Notes
This version took a bit longer, but it's worth the wait! I used a new experimental workflow to get the best out of my previous versions.
The Recipe:
Step 1: Used shootthesound/comfyUI-Realtime-Lora to selectively trim and analyze blocks from v1.0, v1.1, and a bit of v1.15. Highly recommend trying this amazing tool for cleaning up your LoRAs!
Step 2: Merged those trimmed results using my new Easy LoRA Merger tool. It helped me mix the styles without the "burnt" look you sometimes get with standard mergers.
Want to try the merger yourself? It’s currently in the queue for the ComfyUI Manager, but you can grab it here early: 👉 https://github.com/Terpentinas/EasyLoRAMerger
v1.1 - slight improvement - still not there.
v1.0 - My First LoRA for klein‑4B, still in testing phase – feel free to test and give feedback
it focused on breasts, nipple, areola.
Good - works with other Lora tested.
Bad - clothing interaction.
Description
FAQ
Comments (9)
Thanks a lot for training this LoRA. 4b needs love too. I tested it, it works great with character LoRAs too.
Agree 4b is amazing. Thanks for the feedback
@terpentinas totally, 4B is actually nicer than 9B IMO and not just because of speed.
thanks! it works, but... looks like training images contained bodies and faces? it keeps changing a lot of the character, and the skin too, so I can only use it as an inpaint for now.
Yes dataset is near 600 images it have full bodies, various poses, ethnicities.
I think I need refine my dataset - maybe to remove some images and add some sketches and drawings to make it more flexible.
Can you explain then problem happens, what method you use to preserve character? is it overpower your Lora, or you creating something not realistic - like anime.
@terpentinas When you train a LoRA for a concept not a character (your LoRA is a concept not a character) then your dataset should contain only bodies without heads. This way the model will not learn faces and will only learn bodies. Then users can apply their characters' LoRAs to your concept LoRA.
@Learning2025 I think you are right, I will try do it in v1.2 - this model seem so different from Illustrious XL - need to update strategies. Also I always used kohya_ss - need time to learn how to use ai-toolkit
@terpentinas You can probably mask the faces, so the training ignores it.
@terpentinas Definitely look at Ai toolkit, much better control. And watch the official yt tutorial, it is very helpful.