alembics / disco-diffusion

Other
7.47k stars 1.13k forks source link

Controlling object uniformity #125

Open doit-ceo opened 2 years ago

doit-ceo commented 2 years ago

While the framework is great in generating complex landscape, desert, trees, building, etc. I'm looking to control how it generate single object like a car and limit the craziness of the output.

I liked the background, it's even realistic but the car is way unrealistic, how can I fix that ?

car1

aletts commented 2 years ago

For maximum cohesiveness, it can be a good idea to match the output resolution to the diffusion model's trained image size. So in this case, you'd want to experiment with setting the image output size to 512x512. Of course this is less interesting though.. and it might not work really well anyway. Providing an init_image from another source can be a good idea. Some people like to create init images in Midjourney or a latent diffusion model (exclusively, since I believe Midjourney uses some latent diffusion too). That'll give Disco Diffusion some structure to build upon.