The recent developments in generative modeling have seen a significant shift towards more sophisticated and efficient modeling techniques. Researchers are increasingly focusing on integrating Bayesian inference, exploring alternative data spaces, and unifying disparate modeling paradigms under common frameworks. Bayesian-based methods, such as Posterior Mean Matching (PMM), are being leveraged to create flexible and adaptive generative models that can handle diverse data modalities. These models refine their approximations iteratively, offering competitive performance in tasks like image and language generation. Additionally, the exploration of frequency-based modeling, exemplified by DCTdiff, has demonstrated superior quality and efficiency in image generation tasks, particularly in high-resolution scenarios. This approach bridges the gap between diffusion and autoregressive models, suggesting new avenues for research in image modeling. Theoretical advancements, such as the unification of diffusion and flow matching under the Generator Matching framework, provide deeper insights into the robustness and construction of novel model classes. Overall, the field is progressing towards more unified, efficient, and theoretically grounded generative models.