Creating LazyTown AI deepfakes on your local computer

[LazyPooky]

You'll be surprised how easy it is to install and use AI software like Stable Diffusion locally on your Windows 10/11 computer. It probably won't take up more than a few 100MB of space. It is the models, which must be downloaded separately, that are quite large, from 2GB to 5GB. How better trained a model is (large model files), the better the output quality and more realistic it is.

For system performance: A graphics card with at least 6GB VRAM is highly recommend.

You can also train your own created model to create perfect images the way you want. It's better to save the money to buy a good computer than spend it on the wizkids cloud websites that let you only do a snippet of what you can do with AI software 🙂

[LazyPooky]

Example:
Stephanie's Christmas outfit by AI, created in 1minute 27sec, on local computer with 6GB VRAM GPU.

Software:
Windows 10
Stable Diffusion v1.6.1
python: 3.10.6
torch: 2.0.1+cu118
xformers: 0.0.20
gradio: 3.41.2
checkpoint: 346e4b5a73

[chuft]

What was the input image for the Christmas themed dress? I am not sure exactly what the program is doing and what you are doing.

It is hard to understand how such a powerful software could be free. That must have taken an enormous amount of work and skill to create, who paid for all that?

Reading about it, it seems it can't create new things without training that home users can't do.

"For example, the training process for waifu-diffusion requires a minimum 30 GB of VRAM, which exceeds the usual resource provided in such consumer GPUs as Nvidia's GeForce 30 series, which has only about 12 GB."

Are people using this to make these nude images? If so, how are they getting around this VRAM problem?

[LazyPooky]

What was the input image for the Christmas themed dress? I am not sure exactly what the program is doing and what you are doing.

Open Stable Diffusion (browser, windows GUI version based on Gradio library )

For inpaint: Choose a trained model, type what you want to be produced [christmas outfit] , upload the picture and erase what you want to replace, and press generate. It's as simple as that.

In this inpaint example I used the latest model realisticVisionV60B1_v60B1InpaintingVAE (2GB). The models are already created and can be downloaded for free, just like the software. I don't know how they made the models and how hard it is to create your own. I only saw it's possible. There are also nudes-models where you can create nudes if you choose that instead of the one I used. I believe all of the online nudes websites have free Stable Diffusion software as back-end, with various front-end scripts which includes subscription.

[chuft]

So you didn't supply the red patterned Christmas image? It just knows the words "Christmas outfit" means make clothes that have a Christmas theme?

[LazyPooky]

So you didn't supply the red patterned Christmas image? It just knows the words "Christmas outfit" means make clothes that have a Christmas theme?

Yes. Every time you press on generate it will produce a different image. It will be more of the same if you give specific instructions.

[chuft]

Well thanks for the education. So even the least graphically savvy user can create such images at home. In that case I am even more sure that before long these fake nudes will become "no big deal" because it will become clear they are everywhere and easily generated. The shock value will wear off once people realize all it takes is some legal downloads and a few words and clicks to make them. The web sites just make it even easier, or possible for people without graphics cards or decent computers, although at the price of (theoretically) the server having a log of who did what.

Interestingly those images imply the AI can tell the image is of a girl.

[chuft]

Now I am curious about what the wiki article meant about having to train a new model in order to make anime characters. I would think there would be such a model already.

[LazyPooky]

Now I am curious about what the wiki article meant about having to train a new model in order to make anime characters. I would think there would be such a model already.

There are already hundreds of models you can use for free. On the other hand, if you want specific outputs it's best to make/train your own model. My knowledge doesn't go that far right now. Would be cool to have LazyTown specific models.

[possessor]

Yes. Every time you press on generate it will produce a different image. It will be more of the same if you give specific instructions.

That could make for some really hilarous shoops..

[chuft]

Actually I think it will kill off the hobby of shooping altogether.

Pooky what happens if you put in "goth outfit and makeup"

[LazyPooky]

I tried to make a short video from a picture with audio: face and mouth animation with AI software (webgui.zip 113MB for windows 10). Works pretty well, but only for photos where the face is facing forward and not too close.

Actually I think it will kill off the hobby of shooping altogether.

Pooky what happens if you put in "goth outfit and makeup"

Makeup is difficult because it reshapes the face. I can do the clothes. I'm not sure what goth is, but I thought it was black mostly?
It all depends what model you use and and what additional information you provide. If you think of that, there are thousands of variations possible.

[boredjedi]

I was ahead of the times. From June 2007 when I was tinkering around with CrazyTalk.
Put a face image on a prebuilt head model and add audio and it would create the animation.

[possessor]

Actually I think it will kill off the hobby of shooping altogether.

Pooky what happens if you put in "goth outfit and makeup"

Good point. Imagine if photoshops become something you don't need to have TALENT to make.
Pooky, what if you put in "rockstar outfit"?

[LazyPooky]

I was ahead of the times. From June 2007 when I was tinkering around with CrazyTalk.
Put a face image on a prebuilt head model and add audio and it would create the animation.

It's a little better after more than 15 years with a simple software AI 🙂