A gigantic tree with purple leaves and a cave in its tree trunk, a sparkling lake inside the tree trunk, planets in the background, seen from the ground , ruins in the distance , fantasy, matte painting
Stable Diffusion is a open and free Ai , I want to say model for generating text prompts to images. Currently you can very easily play with it on Hugging Face. The free one provided by them is limited in processing so you are better off trying to keep prompts short but it still generates some great results albeit small. as you can see from the first image here.
The prompt I used is at the start, it is not the most optimized but I did get an image that I like although I knew it would have an issue with the description, it is good for what it is.
Now if you google a bit then you will find some sites offering Stable Diffusion interfaces to make life simpler, they all trick you for the most part by allowing one or two renders then they charge.
I guess fair enough since it does take quite a bit of resources but if you have a PC able to play some of these tedious triple A games then you have a PC that will run it for you for free.
For the rest of us that want more power than what HuggingFace offers we can try a google collab. I found this tutorial on one such collab which seems pretty slick - https://bytexd.com/best-way-to-run-stable-diffusion-for-free/ .
Specifically the above tutorial runs through this collab: https://colab.research.google.com/github/altryne/sd-webui-colab/blob/main/Stable_Diffusion_WebUi_Altryne.ipynb#scrollTo=vkgBVo5OEpqn , you will need a Huggingface account for steps 1.4 to get an API access key.
I did run into an issue where the Collab kept hitting an error:
and maybe if I have more time will figure it or it will kinda just resolve itself on the next run.
Why I still mention it is because the guy is quite active on the repo so I think if you can get it running that it will be worth the effort of resolving some errors.
A gigantic tree with purple leaves,sparkling lake, waterfall from the clouds, planets in the background, seen from the ground , ruins in the distance , fantasy, matte painting
I changed the prompt a bit for I almost want to say worst results but only because I did not use any generic modifiers like copying a specific artist etc.
I guess it is something and when you have a more full featured version of Stable Diffusion you can select the best one or just one with the best part and use that image as the base for another iteration and inpainting new things.
Entire workflows can be built around just reprocessing things with a little guidance and exclusion rules to produce very specific types of items. Things like game icons, or avatars etc.
One of the toys in a full workflow is upscaling, so maybe you do get a small 200pixel mess of an image. With AI upscaling you can recover quite a lot of details, I used https://bigjpg.com/ now ut in the full Stable Diffusion model it is part of the options to be able to upscale and other cool features to play with.
The first image was so much better. I guess so it goes.
