Modify Diffusion Process to Start from a Reasonable Camera Pose

[nightgoodl]

Thank you for sharing your impressive work.
In the current diffusion model, the initialization phase does not account for a given camera pose. We would like to provide a reasonably good initial camera pose during the diffusion process and then gradually add noise to it. This can help improve the stability of the model's output and reduce the convergence time compared to starting with random noise.
Is there a way to modify the initialization process so that instead of random noise, the model can be initialized with this reasonable camera pose and then gradually add noise during the diffusion steps?