Table of contents
1. What problem to solve
Previously, the resolution improvement used the generative confrontation network. Friends who have trained this thing should know that the successful training of this thing is a very accidental thing, so naturally this is the pain point of the resolution improvement, so there is this article. The diffusion model replaces the article on generating confrontation networks.
2. how to do
When it comes to how to generate diffusion, there are actually two main aspects:
- 1. How to control
- 2. From what
2.1 How to control?
-
1) If you want to increase the resolution, you are actually drawing a new picture completely, so this high-resolution picture also comes from noise.
-
2) But since there is a low-resolution picture, it cannot really come from a noise, and the information of the low-resolution picture needs to be added, because the correlation between low-resolution and high-resolution is very good, so there is no special feature extraction It can be directly added to the layer-by-layer noise recognition. (It is equivalent to directly inputting low-resolution pictures as control information without feature extraction network)
2.2 Where to start?
The author himself stated that in theory, the generation should start from low resolution and finally get high resolution generation, that is, start from noise and generate high resolution pictures under the control of low resolution pictures. I feel that directly migrating from one to another is not what diffusion is good at. If it is easy to do, the author will do it himself, and will not add such a paragraph specifically.
3. Revelation
It is not the best to extract the control information, sometimes it can be directly entered into it, and it is not a fixed mode but also combined with the situation at the time