My Earthsnake Monster got 1st place in the Earth category of the Elemental Extravaganza Contest! As this might get the image slightly more attention than usual (which is next to nothing), let me be a little bit more detailed about the creation process, in case people are curious.
Concept
First I came up with the concept. My goal was to avoid all tropes and try for something at least somewhat creative. It took a while before I got inspiration and settled on a giant snake made of rocks. I guess this is what pushes it to 1st place in the contest.
Visualization
I used Bing Image Creator to quickly generate a bunch of images with various variations and embellishments of the prompt "Enormous snake made of rocks towering over jungle, fantasy."
DALL-E 3 is a great tool to visualize concepts quickly (when your own hardware isn't up to that task). Its massive dataset and flexibility allows it to merge ideas together with simple prompts. You often get exactly what you ask for, or with a nice twist. For example, I also used it to visualize my fire entry with the prompt "Fire elemental mermaid swimming in a magma lake." (I didn't use that one for img2img though.)
Its big weakness is its aggressive censorship, which can be very frustrating, although I've gotten multiple NSFW results out of DALL-E 3 by accident. If your subject isn't humanoid nor female it shouldn't be an issue.
Another big weakness, at least for Bing Image Creator, is its unreliable service, which tends to present itself as if everything gets censored, but I think that in reality the image generation just failed. Trying again later usually works.
Selecting Base Image
My earth snake wasn't censored so I could generate at full speed, getting four images per try. I downloaded everything that caught my fancy and stopped at 34 images. I then used a multi-round pair-wise elimination process to pick my favorite. I ended up with something generated with the simplest prompt. It was the superior composition that made it stand out.
Generating Final Image
DALL-E 3 images can be great, but they're often unrefined, soft, and can suffer from terrible JPEG compression artifacts. And you cannot tweak them, but that's where stable diffusion comes in. I experimented a bit with completely new prompts, IP-Adapter, and various models, until I settled on a simple 1.5x upscaled img2img.
For upscaling I used REALESRGAN_x2Plus followed by a bicubic downscale. I chose that one because I was familiar with it, the image was already quite noisy, and I wanted some sharpening and segmentation to appear, going for a semi-realistic style.
For the model I picked Juggernaut XL - V9+RDPhoto2-Lightning_4S because I wanted to use a versatile model and try out Juggernaut. I went for V9 because it was easier to get started with, and with Lightning for the speed.
The final prompt is "enormous snake made of rocks covered with moss, towering over jungle, blue eyes, fantasy, high resolution, soft light" without negative prompt, at 6 steps with 0.35 denoise. I used DMP++ SDE (SGM uniform) for nice quality (Karras was too incoherent and noisy). I now avoid that sampler because it's slow, running at half speed compared to 2M/3M SDE, defeating the purpose of using Lightning or Hyper.
I posted the image and then basically forgot about it. When the contest was over it started getting some engagement, which was maybe part of the grading process. And then it ended up as 1st in the Earth category! Ask away if you have further questions.
Do keep in mind that I am a nobody and my win was a fluke.
(update)
Here are some more images that I made with similar workflows: