Ok so, diffusion models have been used from generating NEW images. What I'm trying to achieve is modify an already existing image using this process.
Diffusion models are about going from the extremely noised version of an image (litterally just gaussian noise) to the clear denoised version of it, following a markov chain like so:
T -> T-1 -> ... -> t -> t-1 -> ... -> 0
Each step, to go from t to t-1 we use a trained model to extract the "added" noise and remove it from the image.
This is a deterministric process, so given the same starting noise, using the same trained model, we obtain the same final image result.
The problem that I have is with the other way around, going from a clear image to a noise version of it. This process is stochastic if I just generate a new Gaussian noise and add it.
Should I just train a new model (Maybe an autoencoder) to generate a noisy version an input image that the other model can reconstruct?
I'm way too tired for this...
Diffusion models are about going from the extremely noised version of an image (litterally just gaussian noise) to the clear denoised version of it, following a markov chain like so:
T -> T-1 -> ... -> t -> t-1 -> ... -> 0
Each step, to go from t to t-1 we use a trained model to extract the "added" noise and remove it from the image.
This is a deterministric process, so given the same starting noise, using the same trained model, we obtain the same final image result.
The problem that I have is with the other way around, going from a clear image to a noise version of it. This process is stochastic if I just generate a new Gaussian noise and add it.
Should I just train a new model (Maybe an autoencoder) to generate a noisy version an input image that the other model can reconstruct?
I'm way too tired for this...
