- Jan 4, 2022
- 233
- 869
Use words "love" and pleasure" multiple times to gain confidence from AI and convince it , then you try to save the world
(it's works)
(it's works)Here are a few of mine, trying the realistic look, no inpainting or messing around .
Keen to see if you think I'm doing anything dumb with the prompts
Keen to see if you think I'm doing anything dumb with the prompts
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
- Dec 21, 2020
- 1,531
- 3,618
a bit of fan fiction.
You don't have permission to view the spoiler content.
Log in or register now.
Should probably start a new thread if you want to continue but ...
You don't have permission to view the spoiler content.
Log in or register now.
I'd love to find something like this that didn't have an adult filter (and wasn't paid for)
- Jan 24, 2020
- 1,401
- 3,794
coming to bed?..
Honorable mentioning:
As a starting point I used the awesome prompt by no other than the eminent sharlotte from this post .
It has been completely reworked to suit my own purposes.
You don't have permission to view the spoiler content.
Log in or register now.
Honorable mentioning:
As a starting point I used the awesome prompt by no other than the eminent sharlotte from this post .
It has been completely reworked to suit my own purposes.
Last edited:
a pose site.
You must be registered to see the links
their Civitai
You must be registered to see the links
Last edited:
- Sep 15, 2021
- 931
- 3,423
First impressions are that I may have overdone it in some way - I had 122 images in the training set, so went with 55 vectors per token (!)
The best results for her face seemed to be at 3500 steps. I used SD1.5 as the base model, at least initially, to act as a control. It didn't take long, so I may well retry with a more specifically photorealistic model.
The subject was the retired British glamour model and porn star Natasha Vale.
Unfortunately it's probably over-trained as these are with the same seed but varying the prompt as "in a black dress", "wearing a cream sweater", "wearing a leather jacket", "wearing workout clothes" and "wearing a fighter pilot flight suit":
As you can see, it's also giving a lot of chromatic aberration. The teeth are also not great!
Tests with other models yielded similar results, although I did get this beauty:
I'm going to run a test with my usual settings* to get a baseline, then try again to see if I can get your settings working for me (the speed increase makes it well worth spending the effort to get this working right!).
*which will take a loooonnnggg time!
EDIT: I've done some research and it's probably the very high number of vectors. Unfortunately most of the people discussing TIs on Reddit or Github seem to be discussing something closer to Dreambooth. For example, I don't use descriptor text files for each picture for TIs (I do with Dreambooth, but not a TI).
Last edited:
- Mar 17, 2021
- 308
- 1,053
Nice attempt, thank you for reporting it.
I done more test, and i've discovered something interesting using my settings.
I've trained 2 same TI, one on sd1.5 checkpoint and one i my last merged.
On sd1.5 getting chromatic aberration is fast, on TI trained on it max cfg is 7, if you going up you get chromatic aberration .
On my models, i'm able to do cfg 15, with same parameter and same image for train.
The conclusion is the base model do a lot on difference when training.
I trained on "train embedding tab", with [fileword],[name] and do good description of every image for train
Also i started to train an "oily" helper, just for experimenting with train. And i made image my-self on daz3d.
And this is "boom" on train. let me explain:
I've rendered 20 images at 768x768 in different position of a oiled woman,with transparent background. then added a description for every image.
when you train, the image you use is resized at 512x512. but during resize, the alpha channel gone. this is bad.
So i checked "do not resize image" and " Use PNG alpha channel as loss weight ". and really the train go better, and result is good from start.
Also used 16 vector, with 8 word for initialization.
Actually i'm training another at different step, then doing some test before maybe release it.
At the end, i thing using image with alpha is good, so you put only subject in description.
This is my impression, i need to inspect more.
I've to test training face with alpha, this is next test i want to do.
I done more test, and i've discovered something interesting using my settings.
I've trained 2 same TI, one on sd1.5 checkpoint and one i my last merged.
On sd1.5 getting chromatic aberration is fast, on TI trained on it max cfg is 7, if you going up you get chromatic aberration .
On my models, i'm able to do cfg 15, with same parameter and same image for train.
The conclusion is the base model do a lot on difference when training.
I trained on "train embedding tab", with [fileword],[name] and do good description of every image for train
Also i started to train an "oily" helper, just for experimenting with train. And i made image my-self on daz3d.
And this is "boom" on train. let me explain:
I've rendered 20 images at 768x768 in different position of a oiled woman,with transparent background. then added a description for every image.
when you train, the image you use is resized at 512x512. but during resize, the alpha channel gone. this is bad.
So i checked "do not resize image" and " Use PNG alpha channel as loss weight ". and really the train go better, and result is good from start.
Also used 16 vector, with 8 word for initialization.
Actually i'm training another at different step, then doing some test before maybe release it.
At the end, i thing using image with alpha is good, so you put only subject in description.
This is my impression, i need to inspect more.
I've to test training face with alpha, this is next test i want to do.
- May 18, 2018
- 199
- 2,028
Is there a reason why sometimes for the preview render the LORA looks the way it's supposed to, but the final result is not what the LORA should be?
preview render
final result
It's like the LORA just shits itself midway through the render
preview render
final result
It's like the LORA just shits itself midway through the render
- May 18, 2018
- 199
- 2,028
- Jan 24, 2020
- 1,401
- 3,794
First thing is the denoising strength is too high, lower it to ap 0.3 possibly lower you will need to play a bit with this.
Set the Hires steps to 2x of sample steps. You can of course keep it to 40 but I would recommend 20-30 in combination with 40-60 hires steps.
- Jan 24, 2020
- 1,401
- 3,794
Btw. Even if it wasn't the wished result, the image looks awesome.
Last edited:
- Mar 17, 2021
- 308
- 1,053
Ok guy's this is some sample on what i said, simple prompt with non negative and my TI:
All done at cfg 11.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
All prompt is inside, i managed the view,and if you see, i never specify "oil on body, oily, etc..." but just used TI
Also the last image is on TI at 5000 step, seem overtrained.
So 3000 step seem better. also you don't need so much time for train( 15 min at 512x512, 28 min at 768x768 )
I hope you like it.
- Sep 26, 2018
- 3,041
- 3,140
the people that are using AUTOMATIC1111 and it show xformers not enabled there is a fix
open launch.py and find commandline_args = os.environ.get('COMMANDLINE_ARGS', "") and put --xformers into the "" and hit save
open launch.py and find commandline_args = os.environ.get('COMMANDLINE_ARGS', "") and put --xformers into the "" and hit save
- Jan 24, 2020
- 1,401
- 3,794
A more simple and direct way is to just open webui-user.bat in notepad and add the same argument and then save it.
It's a very good tip though, perhaps there are other things we can change or set in this python file.
- Jan 24, 2020
- 1,401
- 3,794
