If you just want to train a lora for a face you can skip the Lora part and use ReActor, I've had some amazing results replicating people.
You must be registered to see the links
- May 18, 2018
- 194
- 1,953
Does anyone know how to upscale an AnimateDiff gif???
Last edited:
You probably just need to get the gif into an image sequence and then upscale each frame.
Hello, quick question, does anyone know a check point that works well for anime/hassaku style sex scenes. I'm struggling to get multiple characters to appear and finding it impossible to create penetrative sex scenes. It's like i'm stuck in a sfw mode(albeit with tits and vaginas). I'm currently leveraging hassaku hentai model, zedmixhentai and chilloutmix.
Are you trying to use LORAs? Civiti has a tone of sex position Loras. You might be able to get something workable, although when it comes to two or more characters interacting SD can be really hit or miss, especially if you ahve a specific idea.
alternately, find a screen grab of a position you want and use img2img and inpaint to make it work.
alternately, find a screen grab of a position you want and use img2img and inpaint to make it work.
Wow, she looks stunning.
But links are dead, could you reshare it?
And maybe you could do her as HS2 chara.
I do comic work
I have a problem when image is a zoom in shot
event if my model and all lora is 2d stylelish
when go face shot head shot with controlnet
it will add more realistic in it
all promt are the same with very difference outcome
I have a problem when image is a zoom in shot
event if my model and all lora is 2d stylelish
when go face shot head shot with controlnet
it will add more realistic in it
all promt are the same with very difference outcome
- Jan 24, 2020
- 1,401
- 3,792
I'm not doing HS2 content anymore unfortunately but I will upload my Lora again.
Kendra
You must be registered to see the links
- Dec 21, 2020
- 1,523
- 3,588
- Sep 15, 2021
- 827
- 3,121
There's an excellent article on
Effectively, an Embedding is "Here's how you draw Cameron Diaz, based on the information you already have" and a LoRA is "Here's the extra data that tells you what Cameron Diaz looks like". Subtly different concepts.
Combining the two is the genius part - you're effectively asking SD to draw Cameron Diaz* as it would normally but with pictures to copy from.
Examples:
*Other attractive female celebrities exist. Apparently.
You must be registered to see the links
about creating Embeddings, but it also includes the advice to combine them with a LoRA/Lycoris/LoCon, but not use the LoRA's trigger word.Effectively, an Embedding is "Here's how you draw Cameron Diaz, based on the information you already have" and a LoRA is "Here's the extra data that tells you what Cameron Diaz looks like". Subtly different concepts.
Combining the two is the genius part - you're effectively asking SD to draw Cameron Diaz* as it would normally but with pictures to copy from.
Examples:
You don't have permission to view the spoiler content.
Log in or register now.
*Other attractive female celebrities exist. Apparently.
- Jan 24, 2020
- 1,401
- 3,792
- Sep 15, 2021
- 827
- 3,121
Indeed! One thing I discovered when collecting the dataset for the TI training - Cameron Diaz has terrible skin! Really, everyone thinks of her as being so much more beautiful than she really is.Buh but.. I prefer the first image..
- Dec 31, 2016
- 316
- 708
Some scenes from The Mask movie might be to blame for some of that...
Given that the movie is from 30 years ago there's a slight chance she's aged as well
- Dec 31, 2016
- 316
- 708
Was the TI and lycoris both trained on the same model and that model was used for generating the images?
It says to use the lora/lycoris without the trigger word, have you seen or done any tests if there's a difference between not including a trigger in the training as well?
- Sep 15, 2021
- 827
- 3,121
The Lycoris and TI were trained on different data sets by different users (the TI's mine*, the Lycoris belongs to the author). The model used for the TI was SD1.5, I believe the same is true of the Lycoris, but can't confirm 100%.
Using the trigger word for the Lycoris doesn't ruin the results too much, but does 'overdo it' significantly. She looks a lot more wrinkly - 'over-detailed' if you will.
The alternative would be to use the trigger word but turn down the Lycoris weight and/or de-emphasise the trigger word.
The suggestion of not including a trigger in the Lycoris/loRA training at all would be an interesting experiment. I don't really have the time to try that though as it would mean training my own LoRA on Ms Diaz.
*and rather overbaked if I'm honest. Nowhere near the author's recommendations and certainly not very flexible.
- Dec 21, 2020
- 1,523
- 3,588
Fun with the IP Adapter - converting movie scenes / game screenshots into their weird replicas.
I have read the glossary but my head is bad so here are my dumb questions say I have about 100-1000 images of a model(woman) should I train lora or embedding with that? And If my input pictures of women has similar feature, eg big tits, small hips, can lora, or embedding understand the general idea/concept I want to get to, I want to create realistic image?
so here are my dumb questions say I have about 100-1000 images of a model(woman) should I train lora or embedding with that? And If my input pictures of women has similar feature, eg big tits, small hips, can lora, or embedding understand the general idea/concept I want to get to, I want to create realistic image?
Last edited:
IMG2IMG: Does it get any better than this?
TLDR: Is creating consistent cartoonized characters (same EXACT art style, same EXACT clothing, hair, facial shape, etc) using img2img possible in a NSFW-supporting version of stable diffusion when using a bunch of photos of the same model? And AS IMPORTANTLY: is it possible to do successful img2img photo-to-cartoon 2+ model hetero porn scenes that don't turn into a freaking living nightmare? Or is the technology just not there yet for AI builds that still allow NSFW porn?
NOTE after TLDR: I did have a much better looking img2img conversion of this without any sort of "missionary" prompt, but the guy was basically eliminated from the scene entirely by the program. His penis... became a sock on the bed smack dab in between her spread legs.
Ok. I know nothing about AI art practically. But have been casually going to a few sites to see what the capabilities are of turning 2+ model hetero porn photos into cartoons via img2img features, to see if it's even worth investing my time downloading and learning Stable Diffusion right now, based on the technology available.
I figured if one of these sites that charges is able to do img2img conversions of a 2+ person sex scene that are consistent and good looking, then it's worth my time right now to learn it better myself off-site, to get exactly the styles and looks I want. Otherwise... maybe better to wait. As you can see from above, it's not looking good. lolol. (This girl was not originally a tran-- or whatever that is. And that... arm? Was not originally a penis. And I did not ask for this. I DID NOT ASK FOR THIS!!
I've seen somewhat good results in single-person scenes converting photos to an img2img cartoon. But before I delve into this, I'm wondering if anyone's had any luck doing so with multi-person scenes. I know there will be a learning curve, of course. But I highly suspect the technology won't be there for another couple of years anyway, even if I move headlong into this, training models myself, etc.
WHAT I'M LOOKING TO DO: The reason why I'm looking for what capabilities img2img has is that I want the characters to be consistent from scene to scene, just converted to 2d art. Same exact hair and clothing as represented by the host image (no minor variations). Same body and face shape. Same art style. Which text to img prompts, as nice as they look sometimes, just can't do. I might be better off just tracing art and working on things in photoshop instead at least for the next 5 years until there's a leap in tech though. I don't know.
Here was one site that seemed to have its stable diffusion down well with a very nice art style. Giving me SOME hope that maybe the technology is there now. Unfortunately, it's a non-porn site that just just works on faces, cropping off everything else. And I assume it's using a new version of stable diffusion that came after that damn porn lock, where-as we're stuck using earlier versions that allow it. Could obviously use a bit of touch-up (her lips, nose, ect), plus the penis isn't exactly cartoonized. But it makes me think there's some hope possibly?
You don't have permission to view the spoiler content.
Log in or register now.
TLDR: Is creating consistent cartoonized characters (same EXACT art style, same EXACT clothing, hair, facial shape, etc) using img2img possible in a NSFW-supporting version of stable diffusion when using a bunch of photos of the same model? And AS IMPORTANTLY: is it possible to do successful img2img photo-to-cartoon 2+ model hetero porn scenes that don't turn into a freaking living nightmare? Or is the technology just not there yet for AI builds that still allow NSFW porn?
NOTE after TLDR: I did have a much better looking img2img conversion of this without any sort of "missionary" prompt, but the guy was basically eliminated from the scene entirely by the program. His penis... became a sock on the bed smack dab in between her spread legs.
Ok. I know nothing about AI art practically. But have been casually going to a few sites to see what the capabilities are of turning 2+ model hetero porn photos into cartoons via img2img features, to see if it's even worth investing my time downloading and learning Stable Diffusion right now, based on the technology available.
I figured if one of these sites that charges is able to do img2img conversions of a 2+ person sex scene that are consistent and good looking, then it's worth my time right now to learn it better myself off-site, to get exactly the styles and looks I want. Otherwise... maybe better to wait. As you can see from above, it's not looking good. lolol. (This girl was not originally a tran-- or whatever that is. And that... arm? Was not originally a penis. And I did not ask for this. I DID NOT ASK FOR THIS!!
I've seen somewhat good results in single-person scenes converting photos to an img2img cartoon. But before I delve into this, I'm wondering if anyone's had any luck doing so with multi-person scenes. I know there will be a learning curve, of course. But I highly suspect the technology won't be there for another couple of years anyway, even if I move headlong into this, training models myself, etc.
WHAT I'M LOOKING TO DO: The reason why I'm looking for what capabilities img2img has is that I want the characters to be consistent from scene to scene, just converted to 2d art. Same exact hair and clothing as represented by the host image (no minor variations). Same body and face shape. Same art style. Which text to img prompts, as nice as they look sometimes, just can't do. I might be better off just tracing art and working on things in photoshop instead at least for the next 5 years until there's a leap in tech though. I don't know.
Here was one site that seemed to have its stable diffusion down well with a very nice art style. Giving me SOME hope that maybe the technology is there now. Unfortunately, it's a non-porn site that just just works on faces, cropping off everything else. And I assume it's using a new version of stable diffusion that came after that damn porn lock, where-as we're stuck using earlier versions that allow it. Could obviously use a bit of touch-up (her lips, nose, ect), plus the penis isn't exactly cartoonized. But it makes me think there's some hope possibly?
Last edited:
- Sep 15, 2021
- 827
- 3,121
OK.I have read the glossary but my head is bad
As with so many things in life, the answer is "It depends". Sorry!
My advice would be: If the model has a relatively ordinary, normal look, reduce the number of images down to the best quality 50 or so (bearing in mind you're going to have to trim them to 512x512) and train an Embedding (aka TI).
If her look is out of the ordinary or you want more than the face to be specific to her e.g. massive tits, bodily features, tattoos, weird hair etc, create a LoRA instead. Use about 30 images for that.
In either case, for a good likeness do NOT select to mirror the images otherwise the face will look too symmetrical. It's often touted in more basic tutorials as a "one neat trick to double the effective number of training images for free!!1!!". It's actually generally a shit idea, unless the subject is inherently perfectly symmetrical.