nevproject/SonicDiffusionV2

SonicDiffusionV2.ckpt was trained on AnythingV3 for 200 epochs of 203 hand captioned images from various artists. Much more customizable in style and context from previous version

Examples: mobian walking in a snowy sidewalk while drinking coffee

mobian studying for a test

mobian, 1girl, baggy black priest clothes, chains, messy blonde hair, glowing eyes, detailed, :D, golden staff, fantasy, golden light, magic circle, windy, forest

((masterpiece, best quality, very detailed eyes,official art, 8k,4k,2k)),intricate details, illustration,(((goddess))),(extremely detailed clothes),((looking up)),((fromabove)), perfect face, detailed face,smile, 1girl,((light blue eyes)),blonde hair, halo, ((glowing light,))upper body,(angel wings),(heavenly background)

SD2.1 version: I also attempted to create a SD2.1 version using the same images, my computer crashed at 30000 steps but the results are still pretty nice. It especially excels at adding styles. Heres some examples:

mobian in ww1 gear, portrait, regal

mobian mage doing research in lab, portrait, regal, painting

Limitations:

Many of the input images were of female mobians and as such the model has a tendency to create females, adding 'boy' should help.

Due to the way I cropped the images, images tend to generate a 'blur' around the edge photos. Next time I might just leave them as is and see what happens.

More images of different scenarios should be added, I tried to add as many as I could but there's a lot of room for improvement.

Help:

If you would like to help with gathering or captioning images please message me at Metal#0407 on discord.