Sign In

Greenmageddon

415
2.5k
2.6k
14
Verified:
Other
Type
Embedding
Stats
1,688
2,552
Reviews
Published
Mar 17, 2023
Base Model
SD 1.5
Trigger Words
<stardeaf-greenmageddon>
Hash
AutoV2
578A0A3A3E
default creator card background decoration
stardeaf's Avatar
stardeaf

This has been trained on macro photos of lichens and moss that I once took. My intention was to use it as style modifier. Created with Invoke-AI. It works for me with some models and with some it does not, so there's room to experiment. I had good results with realistic/general models, e.g. Deliberate and it works great with iCoMix, which is comic style biased merge.

Use style token alone and it will generate fantasy macros of alien looking lichens. The fun is in mixing it with prompt which requires finding right weight balance. So this is not a ready-made, but requires a lot of tinkering. Here's my approach...

First of all, 'quality' keywords are not really needed and do not make much sense. Depth of field concept is imprinted in the style, as the photos all had a lot of it, so you may want to counter it with negative prompt. Other that that, I would just leave negative part alone.

I would start with defining weights for style and the rest of the prompt, which would look like that (using invoke-ai syntax, but I'm sure you will figure how does it translate, same with prompts on example images):

(<stardeaf-greenmageddon>)0.8, (sleepy zombie reading a paper, raining)0.8

Then I would adjust weights until I find a sweet spot that blends both in an interesting way. The right ratio depends on model, prompt and image size, so experiment... For me odd, detailed ratios seem to work best, like 0.55:0.67. When you get just lichen vegetation, take weight out of style side. I also keep cfg and steps low (below 5 for cfg, step in range of 9-18.)

iCoMix model is here:

https://civitai.com/models/16164

Taking the experiment further, I made derivatives of the embedding. These were trained on images created by using the original style with iCoMix model, as I wanted to see how will they perform when used in other, realistic models...

So the other versions are not updates but derivatives, each with it's own mood. Again, use token weighting to add twirliness and alien organic feel to your image.

Have fun.