Image generation model training method and image generation method

Abstract

The invention discloses an image generation model training method. The method comprises the steps of obtaining a training image and text information corresponding to the training image; generating a word vector based on the text information; inputting the word vector into an encoder for processing to generate a first feature vector and a second feature vector; determining the KL divergence betweenthe distribution of the first feature vector and the standard normal distribution; inputting the second feature vector into a generator to obtain a generated image, and respectively inputting the training image and the generated image into a discriminator to obtain a discriminating value of the input image; determining a first loss value between the training image and the generated image and a second loss value between the corresponding discrimination values; and adjusting parameters of the encoder, the generator and the discriminator until the KL divergence, and when the first loss value andthe second loss value meet preset conditions, and obtaining a trained image generation model based on the corresponding encoder and generator.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a training method for an image generation model, a method for generating an image, a computing device and a storage medium. Background technique [0002] At present, in the field of image processing technology, there are many applications of pixel-level image processing, such as ultra-clear image processing, face attribute synthesis, damaged image restoration, and sketch coloring. [0003] On the one hand, it is difficult to handle such pixel-level image generation based on traditional machine learning frameworks. The main reasons are: 1) Traditional machine learning methods rely on artificially designed features, which are often not applicable to image generation. 2) The feature space required for image generation is too large, and traditional machine learning methods cannot express its features. 3) Image generation needs to recover from the feature space to the ...

AI Technical Summary

Problems solved by technology

[0003] On the one hand, it is difficult to handle such pixel-level image generation based on traditional machine learning frameworks.

Mainly due to: 1) Traditional machine learning methods rely on artificially designed features, which are often not applicable to image generation

2) The feature space required for image generation is too large, and traditional machine learning methods cannot express its features

3) Image generation needs to be restored from the feature space to the original image space, which is not competent for traditional machine learning methods

Simply using the generative confrontation network to generate images from text will still get poor generation results

[0005] Therefore, there is a need for a method of generating images that can solve the problems of difficult control and low quality of generated images in text-generated images

Images