Abstract Video rarely reaches viewers in pristine form: it is downscaled, compressed, and stripped of frames as it is stored, streamed, and re-shared. To recover quality, degraded video is increasingly passed through diffusion-based video super-resolution (VSR) or frame interpolation before it is viewed. These restorers are optimized for perceptual quality, yet the perception–distortion tradeoff implies that better-looking output can contain plausible but wrong detail. Whether such restoration repairs or corrupts the facts a video conveys remains untested. We introduce VID-DRIFT, an evaluation that measures semantic, rather than perceptual, change by tracking how video-MLLM visual question answering (VQA) accuracy shifts under a realistic degrade --> restore pipeline. VID-DRIFT applies downscaling, JPEG, H.265, and frame-rate loss at graded severities, restores each degraded clip with diffusion-based VSR or frame interpolation, and evaluates four open-weight video-MLLMs on multiple-choice VQA over the clean, degraded, and restored versions. Restored accuracy declines with degradation severity for every model and restorer and remains below clean in nearly all conditions. Our Semantic Drift Score further shows that restoration often corrupts facts more frequently than it restores them (mean SDS restorer range: 0.52–0.60). Source code for VID-DRIFT generation and evaluation can be found at: SeeVi-Lab/VID-DRIFT.

Abstract Obtaining pixel-level annotations for semantic segmentation is costly and labor-intensive, making semi-supervised learning a practical and scalable alternative for many real-world applications. Rigorous benchmarks are crucial for advancing semi-supervised segmentation with limited labels. However, most existing semi-supervised segmentation benchmarks focus on simple, idealized office-like scenes. In this paper, we present the first post-hurricane segmentation benchmark that evaluates state-of-the-art semi-supervised semantic segmentation models on post-hurricane datasets that present significantly more challenging conditions. Emergency services have begun leveraging computer vision deep learning models to support disaster relief. One key application is post-hurricane semantic segmentation, which helps responders allocate resources and navigate destroyed or flooded terrain. However, these models must perform reliably amid occlusions, debris, class imbalance, and noisy inputs, often with limited labeled data. We present a comprehensive evaluation of existing methods on disaster-related benchmarks, identify key limitations in their performance under these complex conditions, and highlight future research directions for building more robust and generalizable models tailored to post-hurricane scene understanding.

Abstract Artistic style embodies not only aesthetic choices but also the culmination of an artist's creative identity, vision, and years of practice. The emergence of generative AI, especially diffusion-based models, has introduced new risks by making it possible to imitate these styles with little cost. While general-purpose diffusion models already demonstrate strong capabilities in reproducing stylistic patterns, finetuning further strengthens this ability, allowing models to internalize and replicate an artist's style with greater precision and flexibility. This growing risk of style mimicry underscores the need for effective and efficient protection of artistic works. We hypothesize that certain layers in the diffusion model exhibit heightened sensitivity to artistic styles and can be used to learn efficient yet effective style protection. We propose to measure sensitivity through activation strengths of attention layers in response to style and content representations, and assessing their correlations with features extracted from external models. Based on our findings, we introduce an efficient and lightweight protection strategy, StyleProtect, that achieves effective style defense against finetuned diffusion models by updating only selected cross-attention layers. Our experiments utilize a carefully curated artwork dataset based on WikiArt, comprising representative works from 30 artists known for their distinctive and influential styles and cartoon animations from the Anita dataset. The proposed method demonstrates promising performance in safeguarding unique styles of artworks and anime from malicious diffusion customization, while being efficient and maintaining imperceptibility.

Abstract While neural compression has the potential to revolutionize image compression, recent studies have emphasized its ability to introduce subtle artifacts that could alter the image content. Concerned about the impact of such modifications on scientific images, this work explores the potential effects of neural compression on these images, focusing on two critical aspects: semantic understanding and forensic integrity. We use scientific image datasets to assess the performance of neural compression techniques on Visual Question Answering (VQA) and copy-move forgery detection tasks. Our findings indicate that the subtle changes introduced by neural compression do not significantly degrade the performance of state-ofthe-art solutions. In the experiments, neurally compressed images sufficiently preserved the original semantics and forensic traces. Moreover, compared to lossy techniques, e.g., JPEG compression, at similar bit-per-pixel (bpp) rates, neural compression demonstrates a superior ability to preserve both semantic content and forensic traces, even at high compression levels. Our results suggest that neural compression may provide a viable alternative to lossy compression for scientific images.

Abstract The social media-fuelled explosion of fake news and misinformation supported by tampered images has led to growth in the development of models and datasets for image manipulation detection. However, existing detection methods mostly treat media objects in isolation, without considering the impact of specific manipulations on viewer perception. Forensic datasets are usually analyzed based on the manipulation operations and corresponding pixel-based masks, but not on the semantics of the manipulation, i.e., type of scene, objects, and viewers' attention to scene content. The semantics of the manipulation play an important role in spreading misinformation through manipulated images. In an attempt to encourage further development of semantic-aware forensic approaches to understand visual misinformation, we propose a framework to analyze the trends of visual and semantic saliency in popular image manipulation datasets and their impact on detection.