LoRA Model created from 50 public images of Rachel Weisz.
Warning: Might produce nudity if not appropiated prompt.
Supports are welcome : If you like my work, just buy me a beer (https://ko-fi.com/barabasj214).
Special requests: Do you want something in particular or a personal LoRA model? Contact me through my Ko-fi page sending a direct message (https://ko-fi.com/barabasj214).
Description
FAQ
Comments (8)
There's something wrong it's not working for me.
"Couldn't find Lora with name r4ch3lw41sz"
It's installed properly.
It even crashed my Automatic1111.
Just rename the file and use the new name. Place the file in the models/lora directory. It seems to be working for others.
@barabasj214 NOPE.
Maybe It crashed because you used one of the posted images and loaded the metadata un your input, the sampler will load as null. Anyway its not worth to try...
Hoo man!! Usually I do not give a bad reviews but I'm looking at your config that you are over burning your models and I give you a tip:
100_r4ch3lw31sz
On this one Your steps are at 100 multiplied by the 50 images you are using, this is a big mistake, your models will burn and all those weights will be fired to hell XD....
For 50 images down your 100 steps to 10 and rise the number of epoch to 6, check each generated epoch if you are satisfied with your results, Usually with 50 images epoch #3 have good results, but sometimes likeness can be raised on epoch #4 to #7.
Also a variation of multiple resolutions can collide with the precision on the model, try to auto resize your dataset to a fixed resolution, like 512x512 or 712x712 if you want better quality.
Keep the same config but do not over cook your models with 100 steps using more than 50 images. That's crazy!
Also do not use on your trigger tags "a r4ch3lw31sz woman" only use someting like "rachelweisz-yourinitials" or even "r4ch3lw31sz", every checkpoint contains "a woman" and it cause to mix weights of the generic woman with your woman, your trigger tag isolates those weights to be combined with base woman weights, but yours will override and sum the needed in order to achieve likeness.
This is totally crap.
Delete this. Everyone here keeps trying to nicely tell your submissions suck. And just add +1.
There is something really wrong with your SD Install/Workflow. Get it together man. I'm never downloading one of your files again until it gets multiple 5 stars. Total waste of our time.
As you wish.
Instead of simply criticizing, it would be good, with a little empathy, to help me improve my process since I invest a lot of time in it. My Koyha setup is as follows:
LORA type - Standard
Train Batch - 2
Epoch - 1
Caption Extension - .txt
Mixed Precision - fp16
Save Precision - fp16
Seed - 1234
Learning rate - 0.0001
LR Scheduler - Constant
LR Warmup - 0
Optimizer - AdamW8bit
Network Rank - 128
Network Alpha - 128
Enable Buckets - YES (this allows for big pictures to be used in training)
.... in Advanced Configuration ....
-
CLIP SKIP - 2
Max Token Length - 75
Bucket Resolution Steps - 64
Thanks.