The paper explores the potential of diffusion models in unsupervised learning. Such models use a U-Net to predict and remove noise, resulting in the synthesis of high-quality, diverse images. The researchers propose a new attention mechanism for pooling feature maps and introduce DifFormer and DifFeed, new methods for feature fusion and feedback respectively. The study finds that diffusion models outperform GANs and can compete with top unsupervised image representation learning methods for tasks like image classification and semantic segmentation.
Publication date: 29 Nov 2023
Project Page: ?
Paper: https://arxiv.org/pdf/2311.17921
