TL;DR

DPM++ 2M Karras, CFG 7, Recommended weight is 1. Less in more with prompt, keep it short. No negative prompt is best. Tags: blackhole, blackhole in sky(for an on earth image), nebula, nebula in sky, supernova, supernova in sky, galaxy, nebula [in shape of whatever]. Useful modifiers: unique, distant, colorful, vivid. Keep the extra details to a minimum unless you like crazy busy images. Realistic Pony models will make people better than base Pony. If you want the nebula's in the shape of something bump up the weight of this LoRA and try Euler A, Clip Skip 1.

More Info:

DPM++ 2M Karras seems to work best

Euler A also works but you will need to crank up the steps a bit. Higher steps can cause problems so, up to you.

CFG 7 seems to work best. Lower CFG makes everything desaturated and dull. It's an individual preference thing though.

LoRA weight of 1 works well unless you want the detail to focus on something other than the space stuff.

This LoRA seems to hate negative prompts with a passion. Putting things in there has unpredictable results. Thankfully it seems to create nice images without them. That said, I got better results and prompt tolerance with realistic Pony merges.

Less is more with the positive prompts. The more detail you try to include the more detail it will include and it can overwhelm the image. You might have to lower the weight on realism prompts or it will overwhelm with details without any additional detail LoRA or embedding. Again, the Realistic Pony merges mitigate this somewhat.

I have also noticed that going much higher than 30 steps can make the image become disjointed and "busy" with detail, like a surreal painting. That's a plus for generating on Civitai in my opinion. Fewer steps means less buzz spent. This seems to hold true regardless of the Pony model used.

If you have multiple subjects (blackhole AND a person AND a city etc. with loads of details) your best bet is to use it with either ComfyUI or A1111 and whatever method you choose for controlling prompts with multiple subjects. Prompt bleed will be a problem, as it is with most LoRAs, without the additional controlnet stuff. However if you simply use something like: blackhole in the sky, city, man running, it will be fine. It's the extra details that will get all mixed up. The realistic Pony models are somewhat better with this but it can still be an issue if you leave the LoRA weight at 1.

In addition, if you put something like dark and colorful city in your prompt you're going to get mixed results. Instead try, dark city, neon lights to help the model figure out what you mean. Even with Realistic Pony models the prompt bleed for modifiers is an issue.

Tags used to train where: nebula, galaxy, blackhole, supernova, planet in sky, old galaxy, closeup galaxy, view of earth from the moon, closeup blackhole, blackhole in sky, nebula in sky, distant galaxy, supernova in sky, red giant in sky, distant nebula, blackhole in orbit, distant blackhole, giant star in sky, two blackholes

Putting the modifier "unique" in front of your subject seems to help when the LoRA is consistently giving you what you don't want.

Nebula can take the shape of real things now much better. I believe this is due in part to the versatility of Pony. However, Euler A gives better results and the best results have been with a Clip Skip of 1 which Civitai doesn't always allow. If the author of the checkpoint you're using has set the default to 2 it will default to that at generation unless you enter 1 in the text field for Clip Skip and start generation with ctrl+enter. With some checkpoints the slider isn't present at all.

I'm always open to suggestions for how to improve my work and willing to learn new things.