Noam Levy

Noam Levy

Dynamic sparse attention (DSA) reduces the per-token attention bandwidth by restricting computation to a top-k subset of cached key-value (KV) entries, but its token-dependent selection pattern introduces a system-level challenge: the KV working set is fragmented, volatile, and difficult to prefetch, which can translate into poor cache locality and stalled decode throughput. We study these effects by implementing a lightweight indexer for DSA-style selection on multiple open-source backbones and logging per-layer KV indices during autoregressive decoding. Our analysis shows a gap in serving DSA backbones - a potential for a high volume of blocking LL (last level) cache miss events, causing inefficiency; we propose a novel LL cache reservation system to save KV tokens in the LL cache between decode steps, combined with a token-granularity LRU eviction policy, and show on the data we collected how this architecture can benefit serving with DSA implemented on different backbones. Finally, we propose directions for future architectural and algorithmic exploration to improve serving of DSA on modern inference platforms.

Dor Barber, Rony Zatzarinni, Hava Matichin et al.

In cinematography, visual attributes such as color grading, contrast, and brightness are manipulated to reinforce the emotional narrative of a scene. However, conventional Image Signal Processors (ISPs) prioritize scene fidelity, effectively neglecting this expressive dimension. To bring this cinematic capability to real-time camera pipelines during video capture, we introduce EMOVIS (EMotion-Optimized VISual processing). We establish a systematic mapping between a compact set of high-level emotional states (Happy, Calm, Angry, Sad) and low-level ISP controls - including color saturation, local tone mapping, and sharpness - supported by a calibration user study with statistically significant effects across parameters. We propose a control framework that integrates these emotion-driven adjustments into standard ISP hardware without altering the underlying processing stages. Validation via blind A/B testing shows that viewers prefer the emotion-optimized rendering in 87% of trials when the target emotion matches the scene context, indicating that emotion-aligned ISP control improves perceived suitability for expressive visual content.

Noam Elron, Alex Itskovich, Shahar S. Yuval et al.

The state of the art in video super-resolution (SR) are techniques based on deep learning, but they perform poorly on real-world videos (see Figure 1). The reason is that training image-pairs are commonly created by downscaling a high-resolution image to produce a low-resolution counterpart. Deep models are therefore trained to undo downscaling and do not generalize to super-resolving real-world images. Several recent publications present techniques for improving the generalization of learning-based SR, but are all ill-suited for real-time application. We present a novel approach to synthesizing training data by simulating two digital-camera image-capture processes at different scales. Our method produces image-pairs in which both images have properties of natural images. Training an SR model using this data leads to far better generalization to real-world images and videos. In addition, deep video-SR models are characterized by a high operations-per-pixel count, which prohibits their application in real-time. We present an efficient CNN architecture, which enables real-time application of video SR on low-power edge-devices. We split the SR task into two sub-tasks: a control-flow which estimates global properties of the input video and adapts the weights and biases of a processing-CNN that performs the actual processing. Since the process-CNN is tailored to the statistics of the input, its capacity kept low, while retaining effectivity. Also, since video-statistics evolve slowly, the control-flow operates at a much lower rate than the video frame-rate. This reduces the overall computational load by as much as two orders of magnitude. This framework of decoupling the adaptivity of the algorithm from the pixel processing, can be applied in a large family of real-time video enhancement applications, e.g., video denoising, local tone-mapping, stabilization, etc.

Noam Elron, Shahar S. Yuval, Dmitry Rudoy et al.

Many image restoration techniques are highly dependent on the degradation used during training, and their performance declines significantly when applied to slightly different input. Blind and universal techniques attempt to mitigate this by producing a trained model that can adapt to varying conditions. However, blind techniques to date require prior knowledge of the degradation process, and assumptions regarding its parameter-space. In this paper we present the Self-Normalization Side-Chain (SCNC), a novel approach to blind universal restoration in which no prior knowledge of the degradation is needed. This module can be added to any existing CNN topology, and is trained along with the rest of the network in an end-to-end manner. The imaging parameters relevant to the task, as well as their dynamics, are deduced from the variety in the training data. We apply our solution to several image restoration tasks, and demonstrate that the SNSC encodes the degradation-parameters, improving restoration performance.