metadata

license: mit
language:
  - en
metrics:
  - accuracy
pipeline_tag: unconditional-image-generation
tags:
  - art

Photo Regenerator w/Diffusion Denoiser Model

Incrementally adds noise over a series of steps via a noise scheduler, and feeds images through a CNN to learn features/patterns to then "regenerate" the image.

The .ipynb holds the tests and download of the dataset if needed.

Goals:

Denoise a 25-step noisy image (Blurry)
Denoise a 100-step noisy image
Denoise a 500-step noisy image
Denoise a 1000-step noisy image

Dataset Used: COCO 2017