Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.06616

Video Encoders Built on Image Representations

Jusheng Zhang·Wenhao Wang·Longqi Cai·Liangzhe Yuan·Yuxiao Wang·Ming-Hsuan Yang
View original source ↗Read source document ↗Export metadata ↓

Abstract

The design of a video encoder determines when frames begin to interact and which frame-specific visual evidence remains accessible to the language model. Native video pathways couple neighboring frames during visual encoding, whereas image pathways preserve independently computed frame representations but incur a much larger visual-token cost when all image tokens are forwarded. We ask a basic question: whether a compact video encoder can instead be built on image representations. To answer this question, we separate three operations that are often coupled: per-frame representation, cross-frame token allocation, and temporal interaction. A frozen image encoder first produces frame-specific candidates. A question-aware selector then allocates a fixed token budget across frames using relevance, diversity, and cross-frame correspondence, after which a lightweight learned refiner reads neighboring-frame context and writes residual updates only to the retained anchors. This preserves source positions and keeps the visual output at the fixed budget. Across 13 benchmarks and three vision-language backbones, the resulting pathway matches full-image aggregate performance while using only about 28%-35% of its visual tokens. Specifically, on Qwen3-VL-8B, it achieves a 13-benchmark macro-average of 62.75 with 1,535 visual tokens, compared with 62.58 for the full Image pathway at 4,424 tokens and 59.49 for native Conv3D at 2,212 tokens. On Qwen3-VL-32B, it reaches a 13-benchmark macro-average of 66.28, compared with 66.09 for Image, while providing a 2.16x end-to-end speedup. These results show that compact video encoding does not require early temporal mixing: frame-specific evidence can be preserved first, allocated jointly, and temporally contextualized after selection.

Explore related subjects

cs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jusheng Zhang, Wenhao Wang, Longqi Cai, Liangzhe Yuan, Yuxiao Wang, Ming-Hsuan Yang. 2026-10-05. Video Encoders Built on Image Representations. https://arxiv.org/abs/2610.06616

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Event-based Continuous Color Video Decompression from Single Frames

We present ContinuityCam, a novel approach to generate a continuous video from a single static RGB image and an event camera stream. Conventional cameras struggle with high-speed motion capture due to bandwidth and dynamic range limitations. Event cameras are ideal sensors to solve this problem because they encode compressed change information at high temporal resolution. In this work, we tackle the problem of event-based continuous color video decompression, pairing single static color frames and event data to reconstruct temporally continuous videos. Our approach combines continuous long-range motion modeling with a neural synthesis model, enabling frame prediction at arbitrary times within the events. Our method only requires an initial image, thus increasing the robustness to sudden motions, light changes, minimizing the prediction latency, and decreasing bandwidth usage. We also introduce a novel single-lens beamsplitter setup that acquires aligned images and events, and a novel and challenging Event Extreme Decompression Dataset (E2D2) that tests the method in various lighting and motion profiles. We thoroughly evaluate our method by benchmarking color frame reconstruction, outperforming the baseline methods by 3.61 dB in PSNR and by 33% decrease in LPIPS, as well as showing superior results on two downstream tasks.

cs.CV↗
arXiv2026

InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language Models

Large vision-language models (LVLMs) have demonstrated their incredible capability in visual question answering. However, this rich visual interaction also makes LVLMs vulnerable to adversarial examples. In this paper, we formulate a novel and practical targeted attack scenario that the adversary knows only the vision encoder of the victim LVLM, without the knowledge of its prompts and its underlying large language model. This practical setting poses challenges to the cross-prompt and cross-model transferability of targeted adversarial attack, which aims to confuse the LVLM to output a response that is semantically similar to the attacker's chosen target text. To this end, we propose an instruction-tuned targeted attack (dubbed InstructTA) to deliver the targeted adversarial attack on LVLMs with high transferability. Initially, we utilize a public text-to-image generative model to reverse the target response into a target image, and employ GPT-4 to infer a reasonable instruction $\boldsymbol{p}^\prime$ from the target response. We then form a local surrogate model (sharing the same vision encoder with the victim LVLM) to extract instruction-aware features of an adversarial image example and the target image, and minimize the distance between these two features to optimize the adversarial example. To further improve the transferability with instruction tuning, we augment the instruction $\boldsymbol{p}^\prime$ with instructions paraphrased from GPT-4. Extensive experiments on 6 victim LVLMs demonstrate the superiority of our proposed method in targeted attack performance and transferability. In particular, InstructTA achieves an attack success rate of 51.9% on BLIP-2, outperforming the strongest baseline by 10.5%, and consistently yields the highest attack success rates across all evaluated models. The code is available at https://github.com/xunguangwang/InstructTA.

cs.CV↗
arXiv2026

Motif Channel Opened in a White-Box: Stereo Matching via Motif Correlation Graph

Real-world applications of stereo matching, such as autonomous driving, place stringent demands on both safety and accuracy. However, learning-based stereo matching methods inherently suffer from the loss of geometric structures in certain feature channels, creating a bottleneck in achieving precise detail matching. Additionally, these methods lack interpretability due to the black-box nature of deep learning. In this paper, we propose MoCha-V2, a novel learning-based paradigm for stereo matching. MoCha-V2 introduces the Motif Correlation Graph (MCG) to capture recurring textures, which are referred to as ``motifs" within feature channels. These motifs reconstruct geometric structures and are learned in a more interpretable way. Subsequently, we integrate features from multiple frequency domains through wavelet inverse transformation. The resulting motif features are utilized to restore geometric structures in the stereo matching process. Experimental results demonstrate the effectiveness of MoCha-V2. MoCha-V2 achieved 1st place on the Middlebury benchmark at the time of its release. Code is available at https://github.com/ZYangChen/MoCha-Stereo.

cs.CV↗