K. Li

Yufeng Xu, Taiming Lu, Kunjun Li et al.

Pruning promises a shortcut to strong small language models. In this work, we examine this promise by pruning Llama-3.1-8B at pruning ratios of 0.5--0.8 with six methods spanning depth, width, and sparse granularities, under two controlled token-matched settings. (1) With the same training token budget, pruned initialization consistently outperforms random initialization. This shows that the parent model provides a strong starting point, although the advantage narrows as the training token budget grows and as the pruning ratio rises, nearly vanishing at the highest pruning ratio we study. (2) When training from scratch is instead given the full token budget consumed by the whole pipeline, pruning at finer granularities still retains an advantage, while coarser structured pruning can be matched or surpassed. This suggests that the parent model transfers knowledge that additional training tokens alone cannot fully recover, but only at fine granularity. Taken together, our results yield a clear recommendation: with a large pretrained model in hand and a limited training token budget, pruning is better than training from scratch; when the training budget is not limited, training from scratch can be competitive for coarser pruning, so a large pretrained parent is not always necessary.

Kunjun Li, Manoj Gulati, Steven Waskito et al.

Embedded camera systems are ubiquitous, representing the most widely deployed example of a wireless embedded system. They capture a representation of the world - the surroundings illuminated by visible or infrared light. Despite their widespread usage, the architecture of embedded camera systems has remained unchanged, which leads to limitations. They visualize only a tiny portion of the world. Additionally, they are energy-intensive, leading to limited battery lifespan. We present PixelGen, which re-imagines embedded camera systems. Specifically, PixelGen combines sensors, transceivers, and low-resolution image and infrared vision sensors to capture a broader world representation. They are deliberately chosen for their simplicity, low bitrate, and power consumption, culminating in an energy-efficient platform. We show that despite the simplicity, the captured data can be processed using transformer-based image and language models to generate novel representations of the environment. For example, we demonstrate that it can allow the generation of high-definition images, while the camera utilises low-power, low-resolution monochrome cameras. Furthermore, the capabilities of PixelGen extend beyond traditional photography, enabling visualization of phenomena invisible to conventional cameras, such as sound waves. PixelGen can enable numerous novel applications, and we demonstrate that it enables unique visualization of the surroundings that are then projected on extended reality headsets. We believe, PixelGen goes beyond conventional cameras and opens new avenues for research and photography.

Kunjun Li, Zigeng Chen, Cheng-Yen Yang et al.

Visual Autoregressive (VAR) modeling has garnered significant attention for its innovative next-scale prediction approach, which yields substantial improvements in efficiency, scalability, and zero-shot generalization. Nevertheless, the coarse-to-fine methodology inherent in VAR results in exponential growth of the KV cache during inference, causing considerable memory consumption and computational redundancy. To address these bottlenecks, we introduce ScaleKV, a novel KV cache compression framework tailored for VAR architectures. ScaleKV leverages two critical observations: varying cache demands across transformer layers and distinct attention patterns at different scales. Based on these insights, ScaleKV categorizes transformer layers into two functional groups: drafters and refiners. Drafters exhibit dispersed attention across multiple scales, thereby requiring greater cache capacity. Conversely, refiners focus attention on the current token map to process local details, consequently necessitating substantially reduced cache capacity. ScaleKV optimizes the multi-scale inference pipeline by identifying scale-specific drafters and refiners, facilitating differentiated cache management tailored to each scale. Evaluation on the state-of-the-art text-to-image VAR model family, Infinity, demonstrates that our approach effectively reduces the required KV cache memory to 10% while preserving pixel-level fidelity.

Kunjun Li, Cheng-Yen Yang, Hsiang-Wei Huang et al.

This report introduces ReID-SAM, a novel model developed for the SkiTB Challenge that addresses the complexities of tracking skier appearance. Our approach integrates the SAMURAI tracker with a person re-identification (Re-ID) module and advanced post-processing techniques to enhance accuracy in challenging skiing scenarios. We employ an OSNet-based Re-ID model to minimize identity switches and utilize YOLOv11 with Kalman filtering or STARK-based object detection for precise equipment tracking. When evaluated on the SkiTB dataset, ReID-SAM achieved a state-of-the-art F1-score of 0.870, surpassing existing methods across alpine, ski jumping, and freestyle skiing disciplines. These results demonstrate significant advancements in skier tracking accuracy and provide valuable insights for computer vision applications in winter sports.