Deterministic forward process

[djagpal02]

I am trying to run the deterministic forward process of the probability flow ODE underneath stable diffusion. I am using SDXL via diffusers with DPM++.

The standard pipeline runs the backward process (noise to latent to image). I need the forward (image to latent to noise), such that I have deterministic pairs of noise latent and image.

Modelling the stochastic process is not a problem. I have used the add_noise method within the scheduler to add scaled noise. However, this leads to varied noise (as it depends on what noise you sampled). I know I can fix the noise generator seed, and this will make the process deterministic, but I need the exact inverse of ODE that would generate the noise-free latents.

I know that the forward ODE would just be the negative of the backward ODE, but I'm not sure how this would affect the implementation or if there's a hack / easy way to solve this ODE using the existing system.

Any help would be greatly appreciated!