[10~13] Read the following passage and answer the questions.
image text translation
Among artificial intelligence generation models that generate sentences or video audio
The diffusion model shows excellent performance in restoring and converting images.
diffusion
The basic idea is to gradually add noise to the original image.
If you do not add anything and remove the noise frame again, the original image frame
It is possible to restore: noise is unnecessary or unwanted.
It means value. Noise in the original image that contains only the desired values
When added step by step, it becomes a diffuse image containing noise.
several
If you go through the steps, you will eventually have no idea what the original image was.
It becomes an unrecognizable noise image. Added step by step to billions
If you know the noise frame, you can restore the original image frame from the noise image.
You can. Diffusion model is a noise generator, image operator, noise
It consists of a predictor, which operates in the order of the forward diffusion process and the reverse diffusion process.
The net diffusion process adds noise to the image and creates a noise prediction basis.
It is a learning process: in the first stage; Noise mold in noise generator
After creating it, the image operator adds this noise frame to the original image.
Diffusion with Noise
The image is released.
From the next step
The diffusion output from the previous step of the noise created by the noise generator.
Add to the image If you repeat these steps enough, you will end up with
A noisy image is output. The size or distribution of the noise added at this time is
The characteristics, including aspects, are different at each stage.
thus
The noise predictor is
Step by step, the diffusion image frame is input and the noise contained in the image is
Extract the characteristics and express them in numbers.
this
Based on the numbers
predict noise
Inside the noise predictor
These numbers
These are called latent expressions:
The noise predictor finds the latent representation and
Learn how to predict noise frames.
The learning method of the noise predictor is supervised learning among machine learning.
In supervised learning, the correct answer is given to the learning data and output
This is a method of training the model so that the difference in correct answers becomes smaller. noise
When training a predictor
Created by a noise generator
Noise corresponds to the correct answer. The difference between this noise and the predicted noise is
Learn to make the difference smaller.
The despreading process removes noise from a noisy image and
This is the process of restoring the image frame:
noiset
To remove
in the image
step by step
of some nature
Noise
Additive argument
It’s heaven to know
The noise predictor is
It plays a role.
noise image
or middle
If the diffusion image name in the step is entered into the noise predictor, the image
included
By extracting the characteristics of noise,
Two days to find an expression
based on
noiset
predict image
The operator inputs the spread
This noise frame is subtracted from the image and the noise frame at the current stage is raised.
Print the image frame. If you repeat these steps for the diffusion image, you will end up with
Most of the noise is removed, leaving only an image close to the original image.
It happens.
one side,
Learning many types of image frames
of the learned images
potential
If you add a unique number rate to the expression, the image frame will be used in the despreading process.
You can select and create: You can also adjust the values of the latent expression.
of different characteristics
noise is generated
Combining multiple image frames
It is possible to create an image frame that does not exist.
model’s
noise
noise
noise noise