LLaVA 用低成本对齐打通「 视觉编码器 + 大语言模型」起步，LLaVA‑1.5 以更大更干净的数据与高分辨率输入强化理解，LLaVA‑NeXT 拓展 OCR / 数理与多场景 任务；随后分支为 LLaVA‑NeXT‑Video 处理时序视频、多帧推理，及 LLaVA-NeXT-Interleave 支持交替多图文与跨图联推；最终在 LLaVA‑OneVision 汇聚为统一接 口，覆盖图像 / 文档 / 图表 / 多图 / 视频，兼顾效果与效率。 LLaVA 于 2023 年提出，通过低成本对齐高效连接开源视觉编码器与大语言模型，使「 看图 — 理解 — 对话 」的多模态能力在开放生态中得以普及，明显缩小了 与顶级闭源模型的差距，标志着开源多模态范式的重要里程碑。 尽管多模态对齐的接口与架构趋于收敛，真正「 可复现 」的开源路径仍与「 仅开放权重 」存在间距。Qwen2.5‑VL、InternVL3.5 在 OCR、文档理解、数理与跨图 推理上树立高基线，但完整的数据清单、清洗与混合比例，以及对齐 / 采样与训练日程多为部分披露，难以端到端重现。Molmo 以更干净的数据流水线与精细化 设计，在多项评测 ...

Core Insights - The article discusses the transition from perspective vision to panoramic vision, highlighting the "perspective-panorama gap" as a central theme for understanding the challenges and opportunities in this field [6][19]. - It emphasizes the need for a systematic upgrade across data, models, and applications to enhance the usability of panoramic vision technologies [16][19]. Research Background and Motivation - The paper titled "One Flight Over the Gap: A Survey from Perspective to Panoramic Vision" aims to systematically analyze the differences between perspective and panoramic vision, covering over 300 papers and 20 representative tasks [4][19]. - The article provides a comprehensive overview of the challenges faced in panoramic vision, which are categorized into three main gaps: geometric distortion, non-uniform sampling, and boundary continuity [6][9]. Strategies Overview - Four main strategies are identified for adapting tasks to panoramic vision: 1. **Geometric Distortion**: Issues arise when spherical images are projected onto a plane, leading to shape distortion [7]. 2. **Non-uniform Sampling**: Pixel density varies significantly across different regions, affecting resolution [7]. 3. **Boundary Continuity**: The separation of boundaries in 2D images can lead to learning continuity issues [7]. - The article outlines a cross-method comparison to clarify the applicability of different strategies to various tasks [9][15]. Task Toolbox - The article lists over 20 tasks categorized into four main areas: enhancement and assessment, understanding, multi-modal, and generation, along with representative methods and key papers for each task [12][15]. - It highlights the rapid emergence of new paradigms such as diffusion and generative models, particularly in text-to-image/video and novel view synthesis [15]. Future Directions - To transition from "usable" to "user-friendly," advancements must be made in three main areas: data, model paradigms, and downstream applications [16][21]. - Key challenges include: 1. **Data Bottlenecks**: Lack of large-scale, diverse, and high-quality 360° datasets limits general training and reproducible evaluation [21]. 2. **Model Paradigms**: The need for robust models that can adapt from perspective to panoramic vision while maintaining performance across various tasks [21]. 3. **Downstream Applications**: Applications in spatial intelligence, XR, 3D reconstruction, and various industry sectors require effective deployment and compliance [21][22].

Core Insights - The article discusses the core technologies of multimodal large models, focusing on representation learning, translation, alignment, fusion, and collaborative learning [1][2][7][11][14]. Representation Learning - Representation learning is fundamental for multimodal tasks, addressing challenges such as combining heterogeneous data and handling varying noise levels across different modalities [1]. - Prior to the advent of Transformers, different modalities required distinct representation learning models, such as CNNs for computer vision (CV) and LSTMs for natural language processing (NLP) [1]. - The emergence of Transformers has enabled the unification of multiple modalities and cross-modal tasks, leading to a surge in multimodal pre-training models post-2019 [1]. Translation - Cross-modal translation aims to map source modalities to target modalities, such as generating descriptive sentences from images or vice versa [2]. - The use of syntactic templates allows for structured predictions, where specific words are filled in based on detected attributes [2]. - Encoder-decoder architectures are employed to encode source modality data into latent features, which are then decoded to generate the target modality [2]. Alignment - Alignment is crucial in multimodal learning, focusing on establishing correspondences between different data modalities to enhance understanding of complex scenarios [7]. - Explicit alignment involves categorizing instances with multiple components and measuring similarity, utilizing both unsupervised and supervised methods [7][8]. - Implicit alignment leverages latent representations for tasks without strict alignment, improving performance in applications like visual question answering (VQA) and machine translation [8]. Fusion - Fusion combines multimodal data or features for unified analysis and decision-making, enhancing task performance by integrating information from various modalities [11]. - Early fusion merges features at the feature level, while late fusion combines outputs at the decision level, with hybrid fusion incorporating both approaches [11][12]. - The choice of fusion method depends on the task and data, with neural networks becoming a popular approach for multimodal fusion [12]. Collaborative Learning - Collaborative learning utilizes data from one modality to enhance the model of another modality, categorized into parallel, non-parallel, and hybrid methods [14][15]. - Parallel learning requires direct associations between observations from different modalities, while non-parallel learning relies on overlapping categories [15]. - Hybrid methods connect modalities through shared datasets, allowing one modality to influence the training of another, applicable across various tasks [15].