Computer Vision

Researchers propose GeM-NR, a method for editing nonrigid scene changes in multi-view scenarios, leveraging geometry-aware techniques.

Researchers propose a model that learns from a child's egocentric input, combining visual and verbal data to improve learning capabilities. The model is designed to mimic a child's learning process, allowing for continual learning and adaptation.

Researchers propose adapting vision foundation models using existing metadata, reducing the need for labeled data. This approach is presented in a paper on arXiv AI.

Researchers from Carnegie Mellon University have developed REST3D, a method to reconstruct physically stable 3D scenes from a single image. This is achieved through a combination of techniques that ensure visually consistent and interactive 3D scenes.

Researchers Lianghuan Huang et al. submitted a paper to arXiv, formalizing the binding problem in computer vision and pattern recognition. The paper, titled Formalizing the Binding Problem, explores the concept and its implications.

Researchers from various institutions have proposed AdaCodec, a predictive visual code for video MLLMs (Multimodal Large Language Models). The code is designed to improve the performance of video MLLMs by leveraging visual information.

Researchers propose a mixture-density representation for flying-point-free depth estimation, addressing depth ambiguity in computer vision. The approach is described in a paper submitted to arXiv on June 1, 2026.

Researchers propose a monotonic adaptive norm rescaling approach for long-tailed recognition, aiming to improve hyperparameter-friendliness in optimization.

Researchers Kaidi Zhang and Guanxu Zhu proposed a novel view synthesis method using differentiable multiplane images, achieving fast and lightweight results.

Researchers propose TunerDiT, a training-free progressive steering method for multi-event video generation using diffusion transformers. This approach enables efficient video generation without requiring extensive training data.

Researchers found that vision-language models tend to suppress female representations when given ambiguous input, according to a study published on arXiv. The study analyzed the performance of these models on tasks involving gender classification.

Researchers introduce VideoMLA, a low-rank latent KV cache for minute-scale autoregressive video diffusion. This approach aims to improve video generation efficiency.

Researchers introduce GPIC, a dataset of approximately 28 trillion pixels, comprising diverse internet images captioned by a state-of-the-art vision-language model. The dataset is permissively licensed for research and commercial use.

Researchers propose LocateAnything, a vision-language grounding model that uses parallel box decoding for fast and high-quality results, outperforming existing methods in various tasks.

Researchers propose using social gaze consistency to detect AI-generated images, leveraging the fact that humans tend to gaze at specific points in images, which can be inconsistent in AI-generated content.

Researchers proposed a method to improve the capacity of multimodal large language models for subject-driven generation, used in text-to-image synthesis applications.

Researchers proposed a novel method, Channel-wise Vector Quantization, for efficient image processing based on vector quantization, aiming to reduce computational complexity in computer vision tasks.