BeachFoamXL

This is very experimental. All training images were generated with Dall E 3.
I love testing unconventional subjects and mixing concepts.
Dall E 3 is particularly better in understanding long sentences and more convoluted prompts.
But stable diffusion is very good at adapting anything you feed it. And so far, training on DE3 results works really well. But I just cannot afford prompting 80 images for every idea. And you have to prompt for much more, because by my XP, about every 1 out of 6 Dall E 3 result is actually useful.
I go through every image and use photoshop, inpainting and refining if necessary to optimize the dataset as much as possible. But that´s a very tedious task.
That´s why this early version is very limited and more of a proof of concept. But it does really cool stuff. It even works without any trigger words. It puts a lot of things you prompt onto a beach or into sand. It turns a lot of things into frogs, hermit crabs or starfish. That´s because of the very limited dataset. I also adds foamlike structures to stuff and seems to give objects like skulls a distinct look.
I intentionally undertrained it in order to be flexible. When using very few images, overfit is becoming an issue real quick

Trigger words amplify the effects even further:

v1.0:

Main trigger: s3f
other tags used:

hermit crab - those look miserable on all SD and SDXL models. Massively improved now.

stone with googly eyes - was just a funny image. You can put those eyes on other stuff as well.

starfish - creates just that

oyster - same

small moon - yeah. Check out other planets. Saturn especially nice results

frog - I trained on some ugly mofo frogs. They will most likely influence other animals as well. Dialing down the LoRA strength a bit might help

night - will try its best to do a night time shot but training images were limited - Give it some tries.

Clip Skip: 2 CFG: 7 LoRA: 0.6 - 1.2 (reduce CFG in order to prevent oversampling if necessary)

I strongly recommend you to turn all knobs and configs, sampling setting, sampler, scheduler, base clip, refiner clip, etc.
All combinations created interesting stuff.

I will come back to this one in the future.

If you are interested in my models getting trained and evolved, I appretiate any kind of support. That includes just likes, comments or sharing elsewhere.