FVAR: Visual Autoregressive Modeling via Next Focus Prediction

Recent advancements in MeanFlow training have clarified the dynamics between instantaneous and average velocity fields, revealing that effective learning of average velocity relies on the prior establishment of accurate instantaneous velocities. This understanding has led to the design of a new training scheme that accelerates the formation of these velocities, enhancing the overall training process.

Recent research has delved into the representational capabilities of Mamba, a model gaining traction in vision tasks. This study confirms Mamba's relationship with Softmax and Linear Attention, presenting it as a low-rank approximation of Softmax Attention, and introduces a new binary segmentation metric for evaluating activation maps, showcasing Mamba's ability to model long-range dependencies effectively.

A recent study introduces BD-Net, which successfully applies depth-wise convolution in Binary Neural Networks (BNNs) by proposing a 1.58-bit convolution and a pre-BN residual connection to enhance expressiveness and stabilize training. This innovation marks a significant advancement in model compression techniques, achieving a new state-of-the-art performance on ImageNet with MobileNet V1 and outperforming previous methods across various datasets.

A new approach to flow map distillation has been introduced, which eliminates the need for external datasets traditionally used in the sampling process. This method aims to mitigate the risks associated with Teacher-Data Mismatch by relying solely on the prior distribution, ensuring that the teacher's generative capabilities are accurately represented without data dependency.

The newly proposed DeCo framework introduces a frequency-decoupled pixel diffusion method for end-to-end image generation, addressing the inefficiencies of existing models that combine high and low-frequency signal modeling within a single diffusion transformer. This innovation allows for improved training and inference speeds by separating the generation processes of high-frequency details and low-frequency semantics.

A new study introduces Temporal-adaptive Weight Quantization (TaWQ) for Spiking Neural Networks (SNNs), which aims to reduce energy consumption while maintaining accuracy. This method leverages temporal dynamics to allocate ultra-low-bit weights, demonstrating minimal quantization loss of 0.22% on ImageNet and high energy efficiency in extensive experiments.

A new paradigm in continual learning, Annotation-Free Class-Incremental Learning (AFCIL), has been introduced, addressing the challenge of learning from unlabeled data that arrives sequentially. This approach allows systems to adapt to new classes without supervision, marking a significant shift from traditional methods reliant on labeled data.

A new framework called DiP has been introduced to enhance the efficiency of pixel space diffusion models, addressing the trade-off between generation quality and computational efficiency. DiP utilizes a Diffusion Transformer backbone for global structure construction and a lightweight Patch Detailer Head for fine-grained detail restoration, achieving up to 10 times faster inference speeds compared to previous methods.