Palette: Image-to-Image Diffusion Models

This article is about Conditional Diffusion Application , The author puts forward a method based on Conditional Diffusion Of Image-to-Image new baseline. This article tends to apply , It is not innovative in theory .

1. Introduction

The author's work mainly includes three parts :

1. Based on Conditional Diffusion Of Image-to-Image baseline
2. Explored training Diffusion when , L1 and L2 Loss , as well as self-attention The role of layers
3. In four tasks , Colorization, Inpainting, Uncropping, JPEG decompression Tested new baseline, And used Frechet Inception Distance(FID), Inception Score(IS), Classification Accuracy(CA) of a pretrained ResNet-50 classifier, Perceptual Distance(PD) Four indicators to evaluate the performance of the model .

2. Conditional Diffusion

This paper is not very innovative in theory , It's simple . Diffusion In the model Unet The input is number t The result of the step and the noise parameter of the corresponding step , and Conditional Diffusion Only condition information is added to the input x, And the way to deal with it is to add the conditions x With the original input concatenate get up .

The form of loss function is the same as Diffusion The only difference is that more condition information is entered x.

The improvement made by the author is mainly to DDPM Of Unet Improvements have been made. , In this part, the source code is not given , I don't know much about what I did .

3. experimental result

The author tested four indicators on four tasks , Please refer to the original text for details of the experiment .
It is worth mentioning that , The author tested L1 and L2 Effect of loss on experimental results .

The conclusion is that : L1 and L2 The sampling quality of the loss results is similar , however L2 The results will be more diverse , and L1 The result of the loss is relatively conservative .

Here is an experimental detail , That is to say, when the author is training , A series of noise parameters of the forward process are (1e-6, 0.01) 2000 Step , The sampling process uses (1e-4, 0.09) 1000 Step . This detail can be seen , Training and sampling can have different steps , Different parameters , This is a discovery I haven't seen before .