simple phrases work well
no negative needed (it works better without them)
your loras probably will not work here... at all..
prefers high res fix, helps with grain
This was the first SD 1.5 model that I trained while back. It was tailored to generate images for a bizarre blog my wife and I have called Galactic Mirror. We have become less active on that, so I figure now would be a good time to share it in case someone else can get some use out of it. It was trained on over 2k images that we spent months searching for, hand picking, editing, and captioning to make it so the model would only generate the most bizarre and strange images in a retro futuristic style that we love. We seriously spent way too much time making this. I cannot express that enough.
It does not need, nor like, very long prompts. Simple prompts like "potato" will render the most bizarre imaginings of a potato, without anything else needed. In general, you will get extreme variations from one generation to the next. It is great at animal human hybrids, insects, foods, and really anything. Type in your favorite character and it will make them bizarre.
This will not function like most other models, so do not expect it to. We overtrained the hell out of it so it barely remembers anything other than the completly bizarre. My wife and I have spent hours typing in simple ideas into this thing to see the amazing creations it comes up with. I hope some of you will get to enjoy it as much as we have.
Description
FAQ
Comments (11)
man, I really like your work. Seriously I can play with this all day. Super cool stuff!
So cool!!!!!!
So glad for your work on this, can't wait to see some weird aliens and monsters! There is seriously not enough of this kind of thing around. I have almost 200 photo and CG/anime models each, but my horror folder is less than a dozen.
This is seriously awesome, the pics on the blog are amazing as well. Surprised we haven't seen them here, or if we have they didn't get enough love. The writing on the blog, is that ai as well?
Thank you. Yes the writing is mostly Chat GPT 4.0, but We did a lot of cleanup and fighting with it to make it have the "voice" we wanted
At least, you don't need to input negative prompts :D
gotta be my favourite rn! ᕕ( ᐛ )ᕗ
I'm intriguied. . . What kind of system did you use to file your images? Did you tag them or put them in folders?
I have a little node / react ui I made to help tagg everything. It is currently ugly as hell, but it works. I just use comma delimited text files with the same file name. They work with my trainer and sd-scripts.
Basic flow was:
- Find images we love with various themes to try to keep it balanced, dump them to a folder on a network. Some nights we would look for people, buildings, objects, whatever we felt was lacking in the dataset.
- Manually preprocess with cropping, enlarging, color editing or whatever is needed
- Run blip 2 captioning on the group. Just captions, no tags
- Manually remove trigger terms we wanted to normalize. For example, any mentions of weird, bizarre, strange, etc were removed. Manual tags were added when needed.
- Then just over trained for a few days with a high prompt dropout to burn the concepts in.
Appreciate you taking the time to explain your process. Organizing image sets is complicated because a single image can usually be filed in several places depending on the classification you assign to it. Manual tagging is such a chore - but an essential step that simply must be worked out and worked through. :-)
this is just mind blowing