Guizmus' test models

SDArt v2 (13 PoW)

• training: 593 pictures, 59 300 steps on 1.5e-6 polynomial LR in EveryDream2, base 1.5 +VAE
• Activation token : none, trained as artistic model. Themes trained :
  • lots of medium and generic concepts : anime artwork, painting, artwork, 3D render, realistic photo, portrait, close up, ...
  • PoW Alice in Diffusionland : Alice, cat, rabbit, hatter, psychedelic
  • PoW Catstravaganza : Cat. in lots of situations, with clothes, ...
  • PoW Cosmic Horrors : Horror, tentacles, mist, black and white, cows, cthulhu, glowing eyes
  • PoW Abstract Realities : Broken, abstract, colorful
  • PoW Elemental Flowers : ice, fire, water, earth, flowers, elements
  • PoW Futuristic Fashion : futuristic clothes, tech, neon lights, cyberpunk
  • PoW Underwater Worlds : underwater, mermaid, coral, sea
  • Also included but auto captionned, there is less focus on specific tokens like the previous ones : Imaginary friend, Never ending loop, someone, something, somewhere, synesthesia.
• Model link
• token list
• dataset link

SDArt mix v1 (7 PoW)

• training: 329 pictures, 32 900 steps on 1.5e-6 polynomial LR in EveryDream2, base 1.5 +VAE
• Activation token : SDArt also look at the token list pdf
• Model link
• token list
• dataset link

Bad Cosplay man v1

• training: 42 pictures, 24 000 steps on 1e-6 polynomial LR in EveryDream
• Activation token : Bad cosplay man as
• Model link
• dataset link

Rick Roll v1

• training: 20 pictures, 2400 steps on 1e-6 polynomial LR with 8 bit adam
• Activation token : Rick Roll
• Model link

TMNT v1.0

• training: 11 pictures, 1000 steps on 5e-6 polynomial LR with 8 bit adam
• Activation token : TMNT person
• Model link

Sid v1.0

• training: 10 pictures, 2500 steps on 1e-6 polynomial LR
• Activation token : Sid
• Model link

Kirby v1.0

• training: 10 pictures, 2500 steps on 1e-6 polynomial LR
• Activation token : Kirby
• Model link

Simsbooth v1.0

• training: 122 pictures, 18k steps on 1e-6 polynomial LR
• Activation token : Simsbooth person
• Model link

Hogleg Style v2.4

• training: 5k steps lr 1e-6, 122 samples, then 2k more steps on a lower lr and a rebalanced dataset
• Activation token : HogLeg Style
• Model link

Illusion v1.0

• training: 1k2 steps, 12 samples
• Activation token : Illusion drawing
• Model link

NikeAir v1.0

• training: 1k2 steps, 10 samples
• Activation token : NikeAir Shoes
• Model link