Diffusion

How it works

Algorithm that can turn a picture with bunch of noise into an image described.

Diffusion models are trained by taking a completed HD image then adding random noise to the picture a bit at a time until image is completely random mosaic.

  • We call this downsampling.

Then you train the AI to remove the noise from imperceptible mosaic image to the original HD image a bit by bit at a time.

  • We call this upsampling.
  • inverse of downsampling step.

We train the model to predict the noise added at each step of downsampling.

  • You train by telling it whether it successfully added the noise (predictively) or not.

If the model can predict the noise added at each step, it can take out the noise by reversing the equation.

  • Effectively, generating an image from a random noise.
// training equation
Original_HD_Img + Noise% = Partially_Visible_Img

// reversed equation
PartiallyVisible_Img - Noise% = Original_HD_Img (generated image)