I dont know the difference from GAN.

[AlbertLin0]

My understanding: the picture above has a network $I_\phi $, which is the author's aim to play a role of one-shot inference. The network is trained with variational method and the lower bound is the picture follow.

But, if you only take notice to loss function and network $I_\phi$, there is two questions:

• why dont use truth as $p(x)$, perhaps the diffusion outputs Gaussian means as $p(x)$..
• what's the difference from GAN? freeze the Discriminator to train generartor...

---

THANKS

[sojinleeme]

Thank you for your attention:)

• Q1: Why not use p(x)? A1: We cannot explicitly compute $$p(x)$$, so we approximate it using a pre-trained diffusion model. The diffusion model provides the derivative of $$log (p(x))$$.

• Q2: How does this differ from GANs? A2: Our one-step network resembles GANs in terms of one-step sampling. However, our loss function is fundamentally different from the GAN approach. Unlike GANs, we use the score distillation loss derived from variational inference.

Please refer to our paper for a detailed explanation. We have thoroughly described our main objectives from the perspective of variational inference.