Note: Attached images are quite large (~5MB) and can take a few seconds to open.

While current AI models are quite good in being able to do images with basic messes, a little generic slime, pies, etc., they miss out on some of the finer details and degree of control that comes with the base Stable Diffusion models. By training custom models on high-end consumer hardware, we are effectively able to teach the model a given concept.

In this case, that concept is "maple syrup".

I've spent about 2 weeks going through about a hundred different variations of finetuning the Stable Diffusion XL model (with each run taking 1-4 hours) and have developed something I think is quite good. In comparison to many of the other images posted here, these are able to show a wide variety of clothing, positions, locations, etc. I don't need to fight the model and work around it to get something reasonable. Almost every image produced by the model is great right out of the box. Additionally, since you have more control over the generation, it is possible to utilize the same generation seed to incorporate different views, positions, outfits, within the same scene.

The process was quite difficult and really stretched my graphics card to its limit, but I also learned a ton about training Stable Diffusion models and build and caption effective datasets.

If you would like to produce similar images yourself, I have posted the LoRA which I extracted from the model on CivitAI here:

https://civitai.com/posts/2023554

Fooocus is a great option for using the model locally (Note: You will need a decent Nvidia GPU with at least 6GB of VRAM):

https://github.com/lllyasviel/Fooocus
https://www.youtube.com/watch?v=zIhODzEVZqg

If you are new to running Stable Diffusion models locally, I would highly recommend reading this following introduction, which will teach you everything you need to know:

https://education.civitai.com/civitais-100-beginners-guide-to-generative-ai-art/

I've also included a variety of example images that the model outputs, which have been created at about a 2k resolution and a variety of aspect ratios. Theoretically, there's no reason 4k resolution images could not be an option, besides being massive.

Let me know if there's anything specific you'd like an image of. I'd be happy to produce more. Use your imagination.

Next up, I hope to produce a series of models with concepts like "custard", "baked beans", etc. Let me know if you have any ideas for others as well!

These pictures are all amazing. So realistic!

I have to say the pace at which synthetic imagery is progressing is very impressive. The quality of the images you are sharing here are probably the best I seen ever. Thank you for dedicating so much time on pushing AI wam forward. I can already picture in 5-10 years from now those models being avatars we can guide in a video while we watch with apple vision pro 3 and interact with them. keep up the great work. I sadly dont have the hardware and skills, but looking at your progress it makes me want to invest into the equipment and learning.

To answer your question, I would really be looking forward to see how far you can push it
With mud and Clay so many types of mud and opacity.

Thanks so much for sharing!

These look amazing! Thank you so much for working on this. I've just started playing around on Civitai but I've been really struggling to get it to create any mess. I'm so glad you've taken the time and effort to train it.

It's still difficult for AI to draw hands I see.

But most are really nice and realistic!

These are very good images. It's very hard to get good WAM with AI, but these are amongst the best I've seen. Thanks for sharing.

The level of realism on these is astonishing. This is WAM AI art with a more textured, tactile feel to it than any I have ever seen before. I left comments on a number of the individual photos because this is such an impressive batch of work.