π₯π¨ The Simpsons dreambooth model.
This is a fine-tuned Stable Diffusion model based on The Simpsons.
Use asim style in your prompts. The model has some trouble with double pupils and no pupils. Using "cross-eyed" in the negative prompt appears to help?
Sample images:
Samples are made with dynamic prompts, Euler 80 steps @ CFG 12. Negative prompts: watermark, text, signature, cross-eyed
For people / characters: asim style. dramatic beautiful { headshot | portrait } of __person__ {outside { in a garden | in a desert | on a mountain top | at a roman ruin} {at sunrise | at sunset | on an overcast afternoon | in the rain | in the snow | at night} | inside {a fancy living room | on a movie set | a vast empty dark space | a kaleidoscope | an ancient library} with {spotlights | neon lights | soft mood lighting | firefly lights } }. detailed background. For animals: asim style. dramatic closeup national geographic image of a __animal__ in its natural habitat. at {sunrise|sunset|night}. detailed background. asim style. + random prompt from the internet of cool looking structures: steampunk library, tower of babel, tree house, haunted victorian. biomes: asim style. a beautiful {summer | autumn | winter | spring } landscape panorama painting of __biome__ {at sunrise | at sunset | on an overcast afternoon | in the rain | in the snow | at night}
famous places: asim style. a beautiful panorama view of __places__ {at sunrise | at sunset | on a cloudy afternoon | in the rain | covered in snow}. flowers: asim style. a beautiful vase of __flower__ flowers. on a balcony table at { sunrise | sunset | night} . nearby a {bottle of {beer | wine} and a half-empty glass | bowl of fruit}. asim style. + random prompt from the internet. The model mixes well with existing prompts with artists and styles, though not so well with keywords like "photo-realistic."
Based on StableDiffusion 1.5 model (full weights).
Training
Made with automatic1111 webui + d8ahazard dreambooth extension + nitrosocke guide.
100 hand-cut training images. About 70% people, 20% landscapes and 10% animals and objects. Maybe one too many Cletus. Detailed captions were written for each image such as: "A wide shot of a 40-year-old Caucasian man with glasses and a mustache. Dressed in a fishing hat, pink shirt, an olive fishing vest with pockets and brown trousers, sitting in a canoe on a lake. The man is fishing with a red fishing rod. There are trees and mountains in the background at sunset with a few clouds in the sky."
Learning rate was 1.72e-6 for 10,000 steps without prior preservation. Useful tips from the reddit stablediffusion and the discussions on d8ahazard's extension. Notes on training on d8ahazard dreambooth extension discussion.
I am excited to see what people do with this and I would like to improve the eyes, if anyone has suggestions.