Description
FAQ
Comments (9)
Well ... I think it's going to be many broken ankles.
Hey Netapa how did you train your boots to not end up in a janky mess when generating pictures?
I tried training different kinds of boots but I'm not getting it right and got no idea what I'm doing wrong?
Tried with 50 images, high res, 768/768 on 1.5 model, 100 steps per image and settings from Aitrepreneurs video.
Prompting is probably my problem, I actually tried making the Lora twice andalways added a trigger prompt. First time with prompting everything in the image and 2nd time prompting everything but the boots.
Final result is just that I can't get the boots to work. Or maybe it's my pictures, which only show the boots with a bit of the leg, but not actually the person wearing them.
How did you do it? If you got some tips please help!
I used 39 images in 512x512, 100 steps by image too. Mainly like you, some boots with a bit of legs, but also have some with full body, and some with only the boots.
My train base model is deliberate_v2 (Also a 1.5 model).
I don't know where I took my base config, it was in a youtube random video, but here is the JSON config I have used -> https://pastebin.com/QAC7AKrD
I tried to be accurate on the prompt, like examples of my prompt:
- a woman in latex sitting on a bench with black ballet boots
- a woman in black ballet boots and pink tights
@Netapa I'm not an expert (i never trained a LORA), but I suspect your training prompt are wrong!
"a woman in latex sitting on a bench with black ballet boots"
If I remember well, you must describe what the LORA must NOT learn!
Read better the tutorials, and check about this point!
I tryied this LORA and it was a real pain!
The problem is it trying to change the entire image, stressing all the graphic. I think is trained in a wrong way, or is over-cooked... dont know.
Over a generation of 100 attempts, just 1 or 2 decent images is all you can luckyly get.
Also you must find a checkpoint "compatible" with this difficult LORA.
For me, it was Milky-Chicken
https://civitai.com/models/91662/milky-chicken
because is one the most "result guaranteed" I know, (in the semi-realistic field)
I like the "ballet boots". Hoping in a better version, in the future!
Hi there! Your ballet boots are fantastic! Have you given any thought to doing a version for Flux?
I can try, but not sure my CG can handle an efficient Flux Lora training. And I need to build a new dataset as I lost the old one :(.
Done :P
@Netapa Thank you so much! You've done a man's job sir!