Question about the time encoding

[violet-sto]

Hi, I have a question about the time encoding. As DiffPreT is trained like a diffusion model, does the backbone network GearNet explicitly encode the time step?

[Oxer11]

Hi, thanks for the question. This is essentially a difference bewteen diffusion models for generation and pre-training. In pre-training, we do not explicitly encode the time step to accomordate different encoder architectures and keep the consistency between pre-training and fine-tuning, as no noise is introduced during fine-tuning. Instead, we can implicitly encode the noise level (time step) by using a perturbed distance encoder.

SiamDiff/siamdiff/task.py

Line 208 in 2bc9682

self.dist_mlp = layers.MLP(1, [output_dim] * (num_mlp_layer - 1) + [output_dim])