ooooo, so these have turned out to be my favorite ones so far. This time I started with a model that was trained to make more realistic anime style girls and merged it with my trained gunge models.

I love this style, it's not quite real but still definitely does it for me

These are fantastic! It seems to be a limitation of the current generation of models that you can either have photorealism or contextual awareness (ie. the scene looks well-composed, slime with the correct implicit physics etc), but not both. This is a great compromise.

I wish I had enough local compute to have a proper go with SD models. Running everything on colab pro is just too much of a faff for the time being.

Thank you for the kind words.

I'd say the system has no issue with getting a really accurate and photorealistic result if it is desired. For the training data I have given it then the colours and models were all a little 'bubble gum' in terms of them being over stylised to begin with. Which was largely the aesthetic i prefered. If it were trained on someone else's WAM materials and images then the results would probably look quite different. But this is an interesting area for Messmaster to think about at some point in terms of whether the sites images can be scraped for AI generation or not. I currently just used my own photographed WAM images to train with.

For hardware requirements it can pretty much all be done in 8GB of VRAM now if you follow the right tutorials I think. I'm quite a bit more than that but definitely some improvements being made.