Improving Performance of Face Generation for Specific Individuals Based on LoRA

With the development of Artificial Intelligence (AI) technology, an increasing number of technologies have been applied to the AI field, and the technology of AI to generate pictures has also been greatly developed. At the same time, many new problems have arisen. AI-generated pictures can accurately generate graphics for general people, but for some special historical celebrities, due to the lack of stable diffusion training set, it is always impossible to generate clear faces. This article will use the LoRA model to solve the problem of inaccurate face blur generation. By using a special dataset for LoRA model training, the LoRA model can change the content of the generated images by freezing some permissions of the stable Diffusion base model. The resulting image will be evaluated for similarity to the original image. The results demonstrated that the overall similarity of the image faces after using LoRA increased by about 25%. The degree of missing in the training process of LoRA model is gradually decreasing, and the accuracy is gradually increasing. These show that LoRA model training can improve the accuracy of AI image generation and improve the similarity with the human face. This means that due to the high efficiency of LoRA training, the use of LoRA in the future can no longer be just for the change of painting style, but can be applied to a series of more subdivided fields such as face generation, landscape depiction, background replacement, etc.