Greetings!
Those of you who keep track of my Snakes and Stones thread may have noticed at one point that I was attempting to build a Ball Python orientated AI model using my images with Stable Diffusion (v1.5). My original efforts were ineffective due to the same issue that plagues AI generation of human hands (near infinite positions in 3-Dimensional space).
Stable Diffusion (SD) v1.5 is a bit outdated at this point. There have been two notable generations of updates, Ver. XL (SDXL) and then 'updated' to Ver. 3 (SD3). I was planning on building an add-on model for SD3 (as I have a new computer), however, for whatever reason Stability.AI has not released the coding 'weights' for their most advanced version...so I wait (as of 07OCT2024).
Meanwhile, I have been experimenting with SDXL for reptile-type image generation. It performs a lot better than SD1.5 (of which example images are on my profile), however, SDXL (like 1.5) still needs constant manual digital revision to correct issues.
I wanted to share one of the techniques (simplified) I use to create images.
1. I draw a rough sketch of what I want and write a matching prompt:
2. I run it through an AI filter on a 'low noise setting' a few times and pick the closest match to what I want.
3. I make revisions and add any type of colored pixels that I can use the program to reshape.
4. I specifically select and alter anything I don't like, or want improved (via in-painting).
5. I add stuff, manually fix anything that the AI cannot get right, render at high resolution, color correct, and finally watermark.
Ironically, many AI groups consider my method to be 'cheating' (as they are manually shaped/altered digitally).
A lot of people just type in a prompt and generate an image and then post it online. Artistic venues are absolutely flooded with these first generation 'AI' images-which I refer to as 'rough drafts'. These rough drafts compose about 99% of what I find on AI sites. It is mind-boggling, but, there are people who upload hundreds/thousands of these images/day/every day.
I myself am rather tired of seeing rough draft AI images, but, if you would like me to post some of my finished reptile-themed ones (minus the storyboarding)-I will be happy to do so.
Reply With Quote