Forward Diffusion Noising
The training process of diffusion models gradually adds Gaussian noise to an image until only static remains. Sohl-Dickstein and colleagues set it out in 2015, and generation runs the process in reverse.
Spot it
- 01Dissolving into static
- 02Fading structure
- 03Colored Gaussian grain
Why it happens
At each step a small amount of Gaussian noise is mixed in and the image is slightly scaled down, so after many steps the data is indistinguishable from pure noise.
Recreate it in Pixio
Image edit prompt“Show this image partially consumed by forward diffusion noise: heavy colored Gaussian grain covering the picture, large shapes and colors still faintly visible underneath, fine details fully dissolved into static, like halfway through the noising process. Keep the original subject and composition.”
Open PixioUpload your image and pick an image-edit model.