QUESTION ABOUT STAGE3 #39
Description
Hello, what a great job! But I can't understand step 3. Described in the text：
The St noise level at each step is the same, and there is no diffusion process.
Also during the training phase, we need to use the estimated noise intensity to match the time step t and generate training data. This will result in some time t not being trained if the noise levels of the data are similar. During the reasoning stage, errors will occur when seeking clean images step by step.
Did I understand wrongly? It was also said that the training was exactly the same as DDPM. Does the intermediate step of synthetic noise image matching only occur in the reasoning stage? If so, during training, random Gaussian noise is added at each step, and the noise is shuffled only in the inference stage to generate a synthetic noisy image?