The field of generative modeling and visual understanding is witnessing a significant shift towards developing more efficient, scalable, and explainable models. Researchers are exploring novel architectures and training methods to improve the performance of generative classifiers, image and video generation, and visual understanding tasks. A key direction is the integration of adversarial training principles to enhance discriminative robustness and stable generative learning. Another notable trend is the use of self-supervised representations as a latent space for efficient generation, which has shown promising results in class-conditional image generation and text-to-image synthesis. Furthermore, the development of unified tokenizers and pixel flow models is enabling the simultaneous achievement of high-level semantic abstraction and low-level pixel reconstruction, advancing towards the ultimate goal of universal modeling. Noteworthy papers include: Text2Token, which proposes an unsupervised generative framework for text representation learning. UniFlow, which introduces a generic and unified tokenizer for visual understanding and generation. Your VAR Model is Secretly an Efficient and Explainable Generative Classifier, which presents a novel generative classifier built on visual autoregressive modeling. BIGFix, which proposes a method for self-correcting image generation by iteratively refining sampled tokens. Joint Discriminative-Generative Modeling via Dual Adversarial Training, which integrates adversarial training principles for both discriminative robustness and stable generative learning. Adapting Self-Supervised Representations as a Latent Space for Efficient Generation, which introduces a generative modeling framework that represents an image using a single continuous latent token. pi-Flow, which proposes policy-based flow models for few-step generation via imitation distillation.