So I set about creating a few new keywords for the standard stable diffusion model a couple of months back and then training it on some WAM photos (mostly ones I took myself for my old site). These are some of the images it knocked out. I'll get round to trying again with a more recent version of stable diffusion, but I think these are still looking pretty awesome for 100% AI generated images.
So I set about creating a few new keywords for the standard stable diffusion model a couple of months back and then training it on some WAM photos (mostly ones I took myself for my old site). These are some of the images it knocked out. I'll get round to trying again with a more recent version of stable diffusion, but I think these are still looking pretty awesome for 100% AI generated images.
Let me know what you think.
These all look great, can you tell me the technical detail around a couple of things. What do you mean by...
- Creating a few new keywords - Training it on some WAM photos
These all look great, can you tell me the technical detail around a couple of things. What do you mean by...
- Creating a few new keywords - Training it on some WAM photos
Sure. So the stable diffusion model can learn new things by training it on some new images and teaching it a new keyword to describe that new thing. I chose to train it to understand a new keyword that I called 'gunged1'. You can do this either on your own GPU if you have a powerful one with lots of VRAM or you can do it in lots of places using google colab online (eg. https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb). You need to prepare some training images (and concept images) that it will learn from to understand what the new keyword means.
I actually trained this model with two new terms. These were 'gunged1' and 'asiangirl1'. The reason for choosing these names is that it was a term that the AI didn't already know to avoid confusion. So the end result is that I have a custom stable diffusion model now that understands how to create images of someone being gunged. It is still important to craft a good prompt in stable diffusion to get good results and also to use the negative prompt.
Thanks for the great photos. I have always regretted not downloading the sample images back then. Few artists here understand that good lighting is important. Your pictures with colorful slime and great lighting were always my favorite slime pictures.
These all look great, can you tell me the technical detail around a couple of things. What do you mean by...
- Creating a few new keywords - Training it on some WAM photos
Sure. So the stable diffusion model can learn new things by training it on some new images and teaching it a new keyword to describe that new thing. I chose to train it to understand a new keyword that I called 'gunged1'. You can do this either on your own GPU if you have a powerful one with lots of VRAM or you can do it in lots of places using google colab online (eg. https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb). You need to prepare some training images (and concept images) that it will learn from to understand what the new keyword means.
I actually trained this model with two new terms. These were 'gunged1' and 'asiangirl1'. The reason for choosing these names is that it was a term that the AI didn't already know to avoid confusion. So the end result is that I have a custom stable diffusion model now that understands how to create images of someone being gunged. It is still important to craft a good prompt in stable diffusion to get good results and also to use the negative prompt.
Nice, thank you.
I'm just having syntax problems with my python at the moment, then I'll get going when I have my own to play with.
I want to play with GFPGAN, messmaster was wondering when removing face bluring would become a thing, so am going to test it face restoration would do it.
These all look great, can you tell me the technical detail around a couple of things. What do you mean by...
- Creating a few new keywords - Training it on some WAM photos
Sure. So the stable diffusion model can learn new things by training it on some new images and teaching it a new keyword to describe that new thing. I chose to train it to understand a new keyword that I called 'gunged1'. You can do this either on your own GPU if you have a powerful one with lots of VRAM or you can do it in lots of places using google colab online (eg. https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb). You need to prepare some training images (and concept images) that it will learn from to understand what the new keyword means.
I actually trained this model with two new terms. These were 'gunged1' and 'asiangirl1'. The reason for choosing these names is that it was a term that the AI didn't already know to avoid confusion. So the end result is that I have a custom stable diffusion model now that understands how to create images of someone being gunged. It is still important to craft a good prompt in stable diffusion to get good results and also to use the negative prompt.
I have to look into training the model also. How do you do that? Can you point to a good article describing that?
I have to look into training the model also. How do you do that? Can you point to a good article describing that?
If you Google for "training stable diffusion thelastben" you should find quite a bit about training it using the Google colab I linked to before. The tutorials I followed first were for different colabs so I kinda already understood thelastben one when I found it. The advantages of this colab are mostly just speed.
So I set about creating a few new keywords for the standard stable diffusion model a couple of months back and then training it on some WAM photos (mostly ones I took myself for my old site). These are some of the images it knocked out. I'll get round to trying again with a more recent version of stable diffusion, but I think these are still looking pretty awesome for 100% AI generated images.
Let me know what you think.
These are really good! I'm looking to move to SD as midjourney is very limiting with its NSFW filter. Currently don't have a GPU that's up to the job so looking to deploy it on an AWS EC2 instance so I can do like you've done here. I can only imagine what a bit of serious transfer learning and some decent compute could do to the base model if it was trained on a really extensive WAM-related dataset.
Out of interest, are you using the v2 model of SD here as the base model, or 1.5?
This could definitely be done with pie images. I'll give it a try.
This was on stable diffusion 1.5 as it was the most current a couple months back when I did these. I'll do again on 2.1 and see if it is improved at all.
MMasia said: This could definitely be done with pie images. I'll give it a try.
This was on stable diffusion 1.5 as it was the most current a couple months back when I did these. I'll do again on 2.1 and see if it is improved at all.
Another thing to try might be to start from someone else's NSFW-optimised checkpoint - some good ones on https://civitai.com/
Then rather than having to train the two tokens, you can just train the gunged1 token and apply it inside the optimised checkpoint.
Also not sure if you're using regularisation images at the moment, but these are really useful to prevent overfitting the model to the token, which would help it generalise the token to a wider variety of contexts.
Really wish I had enough compute to hand to make serious inroads into it myself, but short of speccing out a new GPU desktop I don't really have many options. AWS G4dn.xlarge instances are pretty expensive to run on an hourly basis, so can't do much more than cheer from the sidelines at the moment.