It depends on what you are running. I have never used perchance or any other cloud service. I would generate this on an unstable model. Unstable means that it has more stuff like this in the fine tune training phase and is more capable of abstract and artistic randomness. If you want very specific stuff, you will probably need advanced workflows. Personally, I tend to collect a lot of LoRAs for my favorite model(s). Then when I keep getting the wrong stuff in some specific area, I will use a LoRA for that specific feature but in negative weight. Like I collect a lot of stuff like fat, ugly, futa, gay, furry and other kinks because I can weight them negatively to stop that behavior in more complex ways that are not possible or as effective in a prompt.

If you have access to ComfyUI, use the RGThree add on node for the Power LoRA Loader. When you have a LoRA loaded in that node, right click on it and select info. That will list all of the training tags that were used in the LoRA. Pick out the ones you want and then copy them using the copy link. This is extremely powerful because you can see if there are tags that conflict with what you are trying to create. You might want to add these to the negative prompt too or find a different LoRA. I might be able to pull this out of Damn Pony or Ride My Pony Real, but I know these models very well and how to coax what I want from them. Look for Metroid Dread based LoRAs I think.

Just so you know... all perchance image generator pages use the same image generator. They just add different text to the prompts, as "styles." But you can get any of the results from any of the generators on any of the other generators--because they all use the same back-end to actually generate the images.