AI Generated Babes

Upscaler said:
These were my best attempts at Sylvia. These were made with a lora. When I tried training with dreambooth I got worst results, still trying to figure that out.

There's far fewer good photos of Roberta than Sylvia. Should still be possible though.
Click to expand...
Wow! Where did you find the lora? did you make it?
very cute,
 
Jovanna_Smith (cam model), Kelly Rowland, Random girl
 

Attachments

  • _image (1).png
    _image (1).png
    2.2 MB · Views: 427
  • dfg.png
    dfg.png
    2.2 MB · Views: 406
  • dvsd.png
    dvsd.png
    1.1 MB · Views: 364
  • Screen Shot 2024-02-11 at 4.41.15 AM.png
    Screen Shot 2024-02-11 at 4.41.15 AM.png
    490.3 KB · Views: 381
  • undefined_image (2).png
    undefined_image (2).png
    2.6 MB · Views: 429
christianAr said:
Wow! Where did you find the lora? did you make it?
very cute,
Click to expand...
I made it. Although now when I want to replicate a model I use dreambooth.
 

Attachments

  • Sylvia (1).png
    Sylvia (1).png
    1.8 MB · Views: 385
  • Sylvia (2).png
    Sylvia (2).png
    1.7 MB · Views: 394
  • Sylvia (3).png
    Sylvia (3).png
    1.9 MB · Views: 358
  • Sylvia (4).png
    Sylvia (4).png
    2.2 MB · Views: 337
  • Sylvia (5).png
    Sylvia (5).png
    2.1 MB · Views: 364
Upscaler said:
I made it. Although now when I want to replicate a model I use dreambooth.
Click to expand...
Now that there are many different ways for faceswaping or ip adapter with controlnet and more, it isn't any longer entirely necessary to train a lora or ti for a character, unless you are trying to recreate the photo style from an era. If you take a look at prompt geeks tutorial about instant id, it's a good example. I typically use loras for the body shape and/or general style and then use faceswaplabs or reactor for the face. Instant id is supposedly more consistent and works better for images that is not photo realism. I have not tried it out though so can't really comment on how it is to work with. Remember that when you are training a lora or ti, you are always training the entire subject and also the environment. There is no such thing as only training a concept or single detail such as an outfit etc. When you are using a lora you will always get the effect of the entire thing. For this reason, it's my opinion that it's best to use a faceswap extension and/or a controlnet model for the face and/or likeness. This gives you more control to fine tune the rest instead of having to wrestle with any unwanted effect of the character lora or ti.
 
Last edited:
MrFoxxx said:
Now that there are many different ways for faceswaping or ip adapter with controlnet and more, it isn't any longer entirely necessary to train a lora or ti for a character, unless you are trying to recreate the photo style from an era. If you take a look at prompt geeks tutorial about instant id, it's a good example. I typically use loras for the body shape and/or general style and then use faceswaplabs or reactor for the face. Instant id is supposedly more consistent and works better for images that is not photo realism. I have not tried it out though so can't really comment on how it is to work with.
Click to expand...
You do realize a person is more than just their face, right? It's also about capturing the features of their body, something faceswap can't do and loras are notoriously poor at.
 
Upscaler said:
You do realize a person is more than just their face, right? It's also about capturing the features of their body, something faceswap can't do and loras are notoriously poor at.
Click to expand...
Yes I agree, it's why I talked about models in controlnet amongst other things. Of course a person is not only a face but it's the key part. The rest is easier to do separate. It's all I said. Lora's and ti's are notoriously inconsistent even with the best possible training. But it all depends on what the project is and what the goal is. Sometimes good enough is fine. Other times when you really want to hone in on the details yes then maybe training a model might be the solution. In most cases it's not needed though. This doesn't take away from the fine work you have done, I'm only talking in general terms and saying for the ones that can't or don't need to train a checkpoint there are other options that can give a nice result.
Btw I tried the instant id thing and at this point it was not worth it for me, since it's for only SDXL currently. The basemodel supports sd1.5 but the extension doesn't for this model at this time but it's going to in the near future I hear.
 
:geek:
 

Attachments

  • 00153-161969083_raw.png
    00153-161969083_raw.png
    1.4 MB · Views: 750
  • 00363-4117657024_raw.png
    00363-4117657024_raw.png
    1.2 MB · Views: 669
  • 00433-3548667112_raw.png
    00433-3548667112_raw.png
    1.5 MB · Views: 654
  • 00644-3407932636_raw.png
    00644-3407932636_raw.png
    516.5 KB · Views: 654
  • 00110-3119128552_raw.png
    00110-3119128552_raw.png
    1 MB · Views: 631
  • 00685-1168621866_raw.png
    00685-1168621866_raw.png
    1.4 MB · Views: 607
  • 01017-448523283_raw.png
    01017-448523283_raw.png
    1.7 MB · Views: 628
  • 00188-728502089_raw.png
    00188-728502089_raw.png
    1 MB · Views: 740
Stability.ai released a new model with a new architecture, Stable Cascade. They claim it gives better prompt adherence and results than the SDXL base model and after fooling around with a bit it seems that's true, although it still struggles with hands and text. While I eventually got it to do simple prompts like show hands, point at the viewer, clap, and give a thumbs up I couldn't get it to do an OK sign or give a thumbs down. Text was better but still had spelling mistakes.

I was more impressed when tried out some wildcards like a woman driving a car made of bread, Gal Gadot with spoons sticking out of her ears, and Trump diving into a pool of cottage cheese.

NSFW generation is, well, bad, but that's to be expected.

Is this going to replace SDXL and become the new base model everyone will train on or will it be like 2.0 and quickly forgotten by the community? No idea honestly.
 

Attachments

  • 20237-3130851668-A photo of a woman clapping her hands.png
    20237-3130851668-A photo of a woman clapping her hands.png
    1.1 MB · Views: 107
  • 20243-84133882-A photo of a gal gadot with spoons sticking out her ears.png
    20243-84133882-A photo of a gal gadot with spoons sticking out her ears.png
    1.3 MB · Views: 127
  • 20240-2055796816-A photo of a woman driving a car made out of bread.png
    20240-2055796816-A photo of a woman driving a car made out of bread.png
    1.4 MB · Views: 115
  • 20251-2324623391-a full body cinematic photo of a nude woman with huge breasts standing on a b...png
    20251-2324623391-a full body cinematic photo of a nude woman with huge breasts standing on a b...png
    1 MB · Views: 156
  • 20230-2931760828-A photo of a woman giving a thumbs up.png
    20230-2931760828-A photo of a woman giving a thumbs up.png
    1.2 MB · Views: 90
  • 20218-3452466064-A photo of a woman pointing at the viewer.png
    20218-3452466064-A photo of a woman pointing at the viewer.png
    1.2 MB · Views: 98
  • 20210-2485672922-A photo of a woman giving 2 thumbs up.png
    20210-2485672922-A photo of a woman giving 2 thumbs up.png
    1.1 MB · Views: 103
  • 20207-2000621705-A photo of a woman holding up both her hands.png
    20207-2000621705-A photo of a woman holding up both her hands.png
    1.1 MB · Views: 115
  • 20199-3710302303-A photo of a woman holding a sign that says _Tits in Tops Forum_.png
    20199-3710302303-A photo of a woman holding a sign that says _Tits in Tops Forum_.png
    1.2 MB · Views: 137
  • 20247-1989915389-A photo of a trump diving into a pool of cottage cheese.png
    20247-1989915389-A photo of a trump diving into a pool of cottage cheese.png
    1.3 MB · Views: 143
Last edited:
Upscaler said:
Stability.ai released a new model with a new architecture, Stable Cascade. They claim it gives better prompt adherence and results than the SDXL base model and after fooling around with a bit it seems that's true, although it still struggles with hands and text. While I eventually got it to do simple prompts like show hands, point at the viewer, clap, and give a thumbs up I couldn't get it to do an OK sign or give a thumbs down. Text was better but still had spelling mistakes.

I was more impressed when tried out some wildcards like a woman driving a car made of bread, Gal Gadot with spoons sticking out of her ears, and Trump diving into a pool of cottage cheese.

NSFW generation is, well, bad, but that's to be expected.

Is this going to replace SDXL and become the new base model everyone will train on or will it be like 2.0 and quickly forgotten by the community? No idea honestly.
Click to expand...
"although it still struggles with hands and text. While I eventually got it to do simple prompts like show hands, point at the viewer, clap, and give a thumbs up I couldn't get it to do an OK sign or give a thumbs down."

I wonder there would be any robust solution for these. Thank you very much for sharing it.
 
Upscaler said:
Stability.ai released a new model with a new architecture, Stable Cascade. They claim it gives better prompt adherence and results than the SDXL base model and after fooling around with a bit it seems that's true, although it still struggles with hands and text. While I eventually got it to do simple prompts like show hands, point at the viewer, clap, and give a thumbs up I couldn't get it to do an OK sign or give a thumbs down. Text was better but still had spelling mistakes.

I was more impressed when tried out some wildcards like a woman driving a car made of bread, Gal Gadot with spoons sticking out of her ears, and Trump diving into a pool of cottage cheese.

NSFW generation is, well, bad, but that's to be expected.

Is this going to replace SDXL and become the new base model everyone will train on or will it be like 2.0 and quickly forgotten by the community? No idea honestly.
Click to expand...
SDXL also promised the sky but unfortunately under delivered. I think SD1.5 is still the best and with LCM there is a faster alternative. Cascade does sound intriguing though. It's going to be interesting to see how it develops. It's only a matter of time before it can do nsfw content.
Thank you for sharing.
 
MrFoxxx said:
SDXL also promised the sky but unfortunately under delivered. I think SD1.5 is still the best and with LCM there is a faster alternative. Cascade does sound intriguing though. It's going to be interesting to see how it develops. It's only a matter of time before it can do nsfw content.
Thank you for sharing.
Click to expand...
What does it take to get good nsfw images? Do the Loras creators have to redo their work for Cascade usage?
 
:emoji_dancer:
 

Attachments

  • 01186-384224478_raw.png
    01186-384224478_raw.png
    515.5 KB · Views: 640
  • 00975-39424928_raw.png
    00975-39424928_raw.png
    794.7 KB · Views: 625
  • 00945-39424922_raw.png
    00945-39424922_raw.png
    2.7 MB · Views: 612
  • 00824-2597880689_raw.png
    00824-2597880689_raw.png
    2.5 MB · Views: 603
  • 01401-679170409_raw.png
    01401-679170409_raw.png
    653.3 KB · Views: 563
  • 01427-2693673452_raw.png
    01427-2693673452_raw.png
    677.3 KB · Views: 559
  • 01664-744083663_raw.png
    01664-744083663_raw.png
    500.3 KB · Views: 587
  • 01986-1632533169_raw.png
    01986-1632533169_raw.png
    1.6 MB · Views: 658
You must log in or register to reply here.

Similar threads

Janka-Whore
AI generated with me, enjoy!
Replies
4
Views
1,123
AlsoNotMe
A
BQretired
AI Generated Virtual Sex Chatting/Sexing Already Here?
Replies
4
Views
737
BQretired
BQretired
K
AI generated images/threads
Replies
3
Views
785
Ptar
Ptar
G
AI generated breasts - help wanted for LoRa
Replies
1
Views
1,232
Janka-Whore
Janka-Whore
BQretired
Riko. AI generated babe - ASIAN
Replies
0
Views
769
BQretired
BQretired
Top