Sign In

Species

232
1.6k
12
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
642
Reviews
Published
Jun 12, 2023
Base Model
SD 1.5
Training
Steps: 72,000
Epochs: 90
Hash
AutoV2
3371007CED

Species

This model is mostly an experiment to see if I could add various concepts that I wanted into the base model. I think it worked well enough so I'm sharing in case anyone else also wants those things

trained at 768x768 using EveryDream2

some folks have asked for LORAs, I've decided I'm not going to bother. I really doubt a lora can reasonably handle this many things. you should find one for the specific species you want if you want a lora

concepts included:

  • "Centaur" and "x taur" (where x is some animal, "deer", "cat", etc), both of these work really well across multiple styles and characters, although you will occasionally get a horse or a mutant

  • "slime" also works very consistently

  • "Klingon" works very well, "bat'leth" is in there but sucks.

  • "Twi'lek" works well, but messes up the head tails about as often as it does arms

  • "Trill" works okay, but tends to add spots in the wrong places, probably needed more training

  • "Ferengi", gets the head shape good, but faces are a bit weird

  • "Asari" works really well, does men fine even though there were none in training data

  • "Satyr" works pretty well, occasionally it thinks "centaur" for no reason I understand

  • "Cardassian" works nicely

  • "Tortle" is good, "ninja tortle" does too ;)

  • it's silly but I added "foodfolk" in

  • "Myconid" generally works, although it wants to just do people with mushroom hats a lot

  • "gorgon" is alright. sometimes it even gets the snake heads kinda okay. wants them to be screaming, so I should have labeled facial expressions better. added a bit of "naga" as well, and the two mix well

  • "giant x" works pretty well given context, like "walking in a tiny city" or "in the mountains"

  • "tiny x" works sometimes, but is inconsistent. definitely needs context. "trapped in a jar" or "on a giant table" work fairly well

  • "salarian" works good

  • "quarian" works good

  • "krogan" works well, but it doesn't really get gender (not a lot of art of krogan women without a veil)

  • "drell" works well

  • "luxan" works pretty good, although it's hyperfocused on the one character it's seen a lot of

  • can do "leviathan pilot" as long as you don't want the body

  • "andorian" is pretty good, forgets the antenna a lot, same for "namekian"

  • "togruta" is good, though it gets the tails messed up sometimes

  • "wookie is good, but it doesn't really get gender there

  • "bolian" is okay, gets it right sometimes but not always. didn't put enough data in there

  • "t-rex" is better trained than it was, gets things right sometimes

  • "gungan" is pretty good, maybe 30% without big mistakes?

  • "drowned folk" came out pretty well imo

  • I added "dullahan" but it really confuses things and rarely gets it right, same for "cyclops"

  • "x elemental" works nicely, try weird things for x ;)

  • "goblin" got a lot of training in here. lots of modifiers too, try these: "pathfinder", "cute", "orc", "half orc". a bunch of more general ones listed below

  • "talaxian" works pretty well

  • "jem'hadar" works nicely in some styles, not all

  • "na'vi" works pretty well but also has a big style bias

  • "kaminoan" is pretty good, although neck length is all over the place

  • added some "human" training to prevent bleed, feel free to use it in the negatives too

  • "garuda" works pretty well

  • "poseable doll" is a fun one, really wants to be a photo though

  • "lizardfolk" and "draconic lizardfolk" are good

  • "viera" is pretty good, although goes full rabbit head more than I'd like

  • "vulcan" works good, some are also tagged "romulan"

  • lots of fantasy things that had some representation got more training and work more reliably now: "naga", "merfolk", "fairy", "angel", "succubus", "minotaur", "mummy", "griffin", "gnoll", "elf", "dryad", "treefolk", "ghost", "harpy", "sphinx", "genie"

  • lots of generic tags were included in training, haven't tested this super well, but all of these should be useable in positive and negative prompts: "blue/red/green/bark/purple skin", "red/glowing eyes", "tail", "devil tail", "pointed ears", "pointed nose", "tusks", "extra arms", "small/hip/dragon/feathered wings", "clothes torn", "holding x", "x on their shoulder", "small/ram/big horns" and probably lots more I'm forgetting

  • "transformation sequence" I added on a lark, and wasn't expecting it to work well. It doesn't exactly work well, but it's better than I expected. "transformation sequence, man into woman" works the best, but "transformation sequence, x into y" is the format and sometimes works for whatever. "3-part" likewise works most consistently, but you can make it a long image and do a larger "x-part" too. you'll almost definitely need to fix some details (especially faces)

feedback welcome! I'll be doing more versions with more species I think, so let me know which ones you want the most