Tag
NexSight introduces verifiable tasks for fine-grained agentic long-video understanding, presenting SightForge, the 310-task SightLine benchmark, and the SightLoop harness; post-training a 35B model on under 6K verified examples nearly doubles its interactive accuracy and outperforms its 397B family model.
VideoNexus 是一个免训练框架,将视频细节感知与证据驱动的规划推理分离,由感知子代理返回紧凑的时间锚定证据,在长视频基准上以更低成本取得更高准确率。
美团开源发布了 LongCat-Video-Avatar 1.6,这是其音频驱动人体视频生成框架的最新升级版本,改进了口型同步、长时序视频一致性(可超过10分钟)以及端到端推理速度。
本文提出 VMem-Bench,一个两阶段诊断基准,系统研究长视频生成中视觉记忆(检索历史画面)对历史一致性与当前故事准确性的权衡影响,发现检索虽提升实体召回率却降低了精确度与禁用实体规避能力。
VideoMaMa++ 提出一种掩码引导的扩散式视频抠图框架,通过分离的 RGB/掩码 token 流以及 preserve-and-refine 可学习嵌入,实现跨片段边界的滑窗推理,从而在长视频中保持时间一致性并提升细节精度。
提出 FramePack-U,一种免训练的常量内存帧打包框架,可将任意长度视频压缩为固定大小的视觉表示,直接插入现成 VLM 推理,在 Long VideoBench、MLVU 和 Video-MME 上显著超越全分辨率基线及现有 token 压缩方法。
美团开源了13.6B参数的LongCat-Video视频生成模型,原生支持文本到视频、图像到视频和视频续写,可在720p/30fps下生成数分钟长视频且无色彩衰减或质量下降,权重采用MIT许可证并允许商业使用。
提出 In-Distribution Forcing (ID-Forcing) 测试时框架,通过自缓存机制避免 KV 条目偏离训练分布,从而将短时视频扩散模型扩展到分钟级长视频生成,并显著缓解漂移问题。
CoEvoWhen proposes a policy-tool coevolution framework that jointly evolves high-level policies and executable media tools from VLM agentic reasoning trajectories to improve ultra-long video temporal grounding without updating model parameters.
KeyRec is a training-free framework for building bounded visual memory in vision-language models, using a recent-observation cache plus a structured event bank to compress long/streaming video context to just 10% of dense visual tokens while setting state-of-the-art results on 13 of 15 streaming and long-video benchmark settings.
JoyAI-Echo is an open-source GitHub repository that provides long-horizon audio-visual generation and omnimodal world models for creating persistent stories and interactive worlds.
NARU is a benchmark for evaluating narrative evolution and cultural reasoning in Japanese extreme long-form videos, constructed through a hierarchical annotation pipeline and native-speaker verification to address gaps in current benchmarks.
SCOUT introduces a recovery-aware agentic framework with an adaptive exploration-exploitation policy for ultra-long egocentric video reasoning, trained via UPS-GRPO, a uncertainty-prioritized RL method with turn-level advantage decomposition. It achieves state-of-the-art results on ultra-long egocentric benchmarks and remains competitive on shorter long-video settings.
Proposes Self Gradient Forcing (SGF), a two-pass training strategy for autoregressive video diffusion models that provides missing supervision for writing useful context memory, enabling strong long-video extrapolation even from short training windows.
This tweet explains LingBot-World-Infinity, an open-weight video generation model that uses a training technique to recover from errors, enabling coherent hour-long videos across multiple scenes.
OPSD-V improves few-step autoregressive video diffusion models by using real long-video data as temporal context during training, providing dense trajectory-level supervision that enhances visual quality and motion dynamics without altering inference mechanisms.
HKU open-sourced ViMax, a multi-agent collaborative long video generation tool that can generate a coherent video with script, storyboard, voiceover, and consistent characters from a single sentence, solving issues like short video fragmentation and character inconsistency. The developer also introduced Taste-Skill, a frontend framework that improves the aesthetics of AI-generated interfaces.
Seedance 2.5 enables consistent AI-generated videos up to 3 minutes in one go, marking a significant leap in AI video generation that could soon lead to TV episodes and feature films.
This paper introduces V-RAGBench, a benchmark for evaluating retrieval-augmented generation over long egocentric videos, and CARVE, a method that adaptively selects retrieval configurations per chunk to improve VideoRAG performance.
OmniMem introduces a modality-aware memory allocation and perturbation-aware selection strategy for streaming audio-visual LLMs, achieving 2-4% absolute accuracy gains over compression baselines on long-video benchmarks.