Method research for generating diversity pictures based on potential diffusion model

A method for generating diversified pictures based on a potential diffusion model comprises the following steps: firstly, adding prompt variable elements into guide words, and updating a dictionary of a word2vec model; then, the guide word is tokenized and converted into an embedded vector. And selecting a sample picture, mapping through an encoder, adding Gaussian noise to perform forward diffusion, combining with the embedded vector, inputting the potential diffusion model to perform reverse diffusion, and generating a picture. The generated pictures are classified by a downstream identifier, and trust scores are calculated to assess picture performance. And if the trust score is lower than the threshold value, adjusting the embedded vector to optimize the picture, and reducing the similarity. And iteratively optimizing until the picture meets the trust score requirement, and finally outputting a diversified picture. According to the method, the diversity and fairness of the picture data set can be effectiv