多媒体：2025–2026 年度综述

An annual survey of the cs.MM arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 120 papers and organised into 44 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• Image Compression and Quality Assessment (20.4%) — Focuses on efficient image compression, quality enhancement, and no-reference quality assessment using diffusion models and neural networks.
• Multimodal Learning and Sentiment Analysis (17.4%) — Explores multimodal representation learning, sentiment analysis, and benchmarking for large language models in diverse multimedia contexts.
• Audio-Visual Deepfake Detection (11.0%) — Addresses deepfake detection, audio-visual representation learning, and lip synchronization for talking head generation.
• Joint Audio-Video Generation (10.5%) — Investigates unified frameworks for joint audio-video generation, video translation, and synchronized audio generation from video.
• Vision-Language Models and Understanding (9.5%) — Centers on multimodal large language models for vision-language understanding, video moderation, and spatial intelligence.
• Multi-Agent Multimodal Systems (9.0%) — Examines multi-agent collaboration, multimodal retrieval-augmented generation, and task-specific prediction in multimedia systems.
• Music Generation and Performance Analysis (6.0%) — Covers music generation, piano transcription, and culturally situated approaches to AI music systems with dataset analysis.
• Cross-Modal Alignment and Retrieval (5.8%) — Focuses on fine-grained cross-modal alignment, attention mechanisms, and retrieval techniques for multimodal understanding.
• Real-Time Multimedia Communication (5.7%) — Addresses real-time video streaming, bandwidth estimation, and low-latency communication systems for multimedia applications.
• Multimodal Emotion Recognition (4.5%) — Specializes in multimodal emotion recognition in conversation, integrating speech, visual, and contextual cues for accurate detection.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.