generation stability

USING NOISES FOR STABLE IMAGE GENERATION IN DIFFUSION MODELS

The foundation of this work lies in the study of the image generation process using diffusion models. The potential for increasing generation stability is demonstrated through an approach that involves using constant noise as a regularizer, with the addition of a mask of random noise constructed around each pixel of the input image within a random radius. The models were implemented using Python and TensorFlow, NumPy, and Keras libraries.