Why learning with backward process is need in diffusion model despite knowing $q(x_{t-1}|x_t)$

Question

I recently learnt about the diffusion model in deep learning, can someone explain to me if we can induce noise to an input data and make it a gaussian like data, why can't we use the same process in reverse without training a network for denoising to deconstruct the gaussian data to the real image?

in other words, in $q(x_{t-1}|x_t) q(x_t) = q(x_{t-1})q(x_{t}|x_{t-1})$ we do have $q(x_t)$ at the very last step which has become an gaussian through introducing noise in iterative manner, and we know what conditional probabilities are, why we try to approximate $q(x_{t-1}|x_t)$, if it just a gaussian.

Answer 1

If you read the paper High-Resolution Image Synthesis with Latent Diffusion Models, you'll notice that the Gaussian is actually $q(x_{t-1}|x_t, x_0)$, whereas $q(x_{t-1}|x_t)$ is intractable. In this case, even if we know the appropriate distribution, we do not know $x_0$ in the denoising step (as that is ultimately what we are trying to predict) and hence we still have to approximate the posterior with some neural network approximation $p_{\theta}(x_{t-1}|x_t)$.