[{"id":"CVPR-2026-0435","venue":"CVPR","year":2026,"title":"VideoCoF: Unified Video Editing with Temporal Reasoner","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Yang_VideoCoF_Unified_Video_Editing_with_Temporal_Reasoner_CVPR_2026_paper.html","doi":"","generation_model":"WAN-14B；推理使用 DMD-LoRA 少步版本","auxiliary_models":"数据构建使用 Qwen-VL-72B；不是推理时必需的外部 reasoner","conditions":"源视频 + 文本编辑指令；模型预测 edit-region reasoning latents，无需用户 mask","training_datasets":"50k 视频编辑训练对／三元组；来源包含 Pexels 的实例级构建数据及 Señorita-2M 筛选子集","evaluation_datasets":"VideoCoF-Bench","evidence_location":"§3.5–4.1，PDF pp.5–6","training_status":"有训练","annotation_notes":"各来源混合数量未在所读段落明确拆分；底座原始预训练数据不归作本论文训练集。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Yang_VideoCoF_Unified_Video_Editing_with_Temporal_Reasoner_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-1584","venue":"CVPR","year":2026,"title":"TV2TV: A Unified Framework for Interleaved Language and Video Generation","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Han_TV2TV_A_Unified_Framework_for_Interleaved_Language_and_Video_Generation_CVPR_2026_paper.html","doi":"","generation_model":"自建 Mixture-of-Transformers：3B/8B text tower + 同规模 video tower；Cosmos-Tokenize1-CV4x8x8-360p VAE","auxiliary_models":"text tower 初始化 Llama-3.2-3B / Llama-3.1-8B","conditions":"meta-prompt + 首帧；交错的文字动作／规划和视频；允许生成中插入文字干预","training_datasets":"CS:GO 95 小时带 controller action 的视频；YT-Temporal-1B 的体育视频，经场景切分、过滤及交错 caption 构建","evaluation_datasets":"游戏与体育视频的质量、控制性人工对比","evidence_location":"§2.1、3.1–3.2、4.1–4.2，PDF pp.3,5–7","training_status":"有训练","annotation_notes":"不是直接微调某个现成 T2V 模型；文本塔预训练与视频训练数据分开记录。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Han_TV2TV_A_Unified_Framework_for_Interleaved_Language_and_Video_Generation_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-1722","venue":"CVPR","year":2026,"title":"DriveLaW: Unifying Planning and Video Generation in a Latent Driving World","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Xia_DriveLaW_Unifying_Planning_and_Video_Generation_in_a_Latent_Driving_CVPR_2026_paper.html","doi":"","generation_model":"LTX-Video 初始化的 2B Video DiT + 133M diffusion planner","auxiliary_models":"无独立 LLM planner；规划器读取 Video DiT latent","conditions":"历史相机帧（规划阶段 past 4 frames）及驾驶状态上下文；预测未来视频与轨迹","training_datasets":"nuScenes + nuPlan；两阶段视频预训练后做 trajectory fine-tuning","evaluation_datasets":"nuScenes 视频生成；NAVSIM 轨迹规划","evidence_location":"§3.5、4.1，PDF pp.5–6","training_status":"有训练","annotation_notes":"NAVSIM 是文中规划评测基准，不能据此替换训练来源 nuPlan。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Xia_DriveLaW_Unifying_Planning_and_Video_Generation_in_a_Latent_Driving_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-2024","venue":"CVPR","year":2026,"title":"Training-free Motion Factorization for Compositional Video Generation","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Wang_Training-free_Motion_Factorization_for_Compositional_Video_Generation_CVPR_2026_paper.html","doi":"","generation_model":"VideoCrafter-v2.0、CogVideoX-2B","auxiliary_models":"Llama-3.3-70B 做 structured motion reasoning；Llama-3.1-8B 为消融","conditions":"文本 prompt → motion graph → 逐帧位置／形状与分解后的运动 guidance","training_datasets":"无新增训练：training-free 推理方法","evaluation_datasets":"CVGBench-m（MSR-VTT 派生）及 CVGBench-p（Panda-70M 派生）","evidence_location":"§4，PDF pp.5–8","training_status":"Training-free","annotation_notes":"MSR-VTT / Panda-70M 在本论文这里用于构建评测，不是本方法训练集。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Wang_Training-free_Motion_Factorization_for_Compositional_Video_Generation_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-2326","venue":"CVPR","year":2026,"title":"MoReGen: Multi-Agent Motion-Reasoning Engine for Code-based Text-to-Video Synthesis","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Bai_MoReGen_Multi-Agent_Motion-Reasoning_Engine_for_Code-based_Text-to-Video_Synthesis_CVPR_2026_paper.html","doi":"","generation_model":"代码驱动的物理模拟与渲染；不以 diffusion video backbone 为主生成器","auxiliary_models":"Qwen2.5-Coder-14B；GPT-5 变体／反馈；Qwen2.5-VL evaluator；GroundedDINO + CoTracker3","conditions":"自然语言物理场景 → 结构化参数 → 可执行模拟代码与轨迹","training_datasets":"1,200 条人工整理 text–specification pairs 对 Qwen2.5-Coder-14B 做 SFT","evaluation_datasets":"MoReSet：1,275 个带轨迹的物理视频；评测集与 SFT pairs 分开","evidence_location":"§3.1–4，PDF pp.4–8","training_status":"有训练","annotation_notes":"模拟／渲染采用 Pymunk 等工具；Blender/Manim 被列为可用引擎，不能一律视作所有实验统一配置。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Bai_MoReGen_Multi-Agent_Motion-Reasoning_Engine_for_Code-based_Text-to-Video_Synthesis_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-2560","venue":"CVPR","year":2026,"title":"SVBench: Evaluation of Video Generation Models on Social Reasoning","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Peng_SVBench_Evaluation_of_Video_Generation_Models_on_Social_Reasoning_CVPR_2026_paper.html","doi":"","generation_model":"被测：Hailuo02-S、Kling2.5-turbo、Sora2-Pro、Veo-3.1、HunyuanVideo、LongCat-Video、LTX-1.0、Wan2.2","auxiliary_models":"Gemini-2.5-Pro 为 VLM judge","conditions":"由社会认知实验转写的文本场景 prompt","training_datasets":"无新增生成模型训练；benchmark 的 agent-based 构建流程为 training-free","evaluation_datasets":"SVBench；心理学实验派生的社会推理任务","evidence_location":"§3–4、Table 2，PDF pp.4–7","training_status":"评测／无新增训练","annotation_notes":"Table 2 列 8 个被测模型；正文模型清单漏写 Wan2.2，以表格为准。底座训练集未由此论文统一披露。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Peng_SVBench_Evaluation_of_Video_Generation_Models_on_Social_Reasoning_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-2680","venue":"CVPR","year":2026,"title":"Active Intelligence in Video Avatars via Closed-loop World Modeling","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/CVPR2026/html/He_Active_Intelligence_in_Video_Avatars_via_Closed-loop_World_Modeling_CVPR_2026_paper.html","doi":"","generation_model":"Wanx2.2 I2V + distilled LoRA（LightX2V）","auxiliary_models":"Gemini-2.5-Flash 同时用于 System 1 / System 2","conditions":"高层任务意图 + 初始／当前视频观测；状态记忆和反馈 → 分步动作 caption","training_datasets":"无任务特定训练；预训练 VLM 的 structured prompting 与现有 I2V 模型","evaluation_datasets":"L-IVA：长程交互 avatar 五类场景","evidence_location":"§3.2、5.1–5.2，PDF pp.4–6","training_status":"Training-free","annotation_notes":"使用已有 distilled LoRA 不等于本文重新训练；其训练来源未在所读实验配置段落披露。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/He_Active_Intelligence_in_Video_Avatars_via_Closed-loop_World_Modeling_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-2789","venue":"CVPR","year":2026,"title":"Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm","reasoning_role":"R2","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Tong_Thinking_with_Video_Video_Generation_as_a_Promising_Multimodal_Reasoning_CVPR_2026_paper.html","doi":"","generation_model":"Sora-2 为主要被研究 video reasoner；Wan-2.5-I2V-preview 用于 prompt-rewriter 消融","auxiliary_models":"GPT-5、Gemini-2.5-Pro、Claude 等为比较的 VLM，非统一生成 pipeline 组件","conditions":"任务图像／题面 + 文本指令；以生成视频帧（部分任务含语音）表达答案","training_datasets":"无新增训练；调用现有模型测试推理能力","evaluation_datasets":"VideoThinkBench：Eyeballing Puzzles、PuzzleVQA、ARC-AGI-2、GSM8K、MATH、MMLU、MMMU 等","evidence_location":"§2–3，PDF pp.3–8","training_status":"评测／无新增训练","annotation_notes":"闭源底座训练数据未披露；评测题集不可写作训练 dataset。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Tong_Thinking_with_Video_Video_Generation_as_a_Promising_Multimodal_Reasoning_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-3054","venue":"CVPR","year":2026,"title":"LAMP: Language-Assisted Motion Planning for Controllable Video Generation","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Kizil_LAMP_Language-Assisted_Motion_Planning_for_Controllable_Video_Generation_CVPR_2026_paper.html","doi":"","generation_model":"轨迹可控的视频生成后端；主文所读部分未明确统一 checkpoint；Wan2.1 明确是 text-only baseline","auxiliary_models":"微调 Qwen2.5-VL 系列 motion planner（由文中引用 [29] 确认系列，参数规模不确定）；GPT-5 thinking 是 in-context 对比","conditions":"自然语言对象／相机运动 → DSL 程序 → 显式 3D bounding-box 与 camera trajectories","training_datasets":"自建程序化 motion-program 数据：文字描述、DSL 及 3D 轨迹配对；训练 planner","evaluation_datasets":"DataDoP、ET 等轨迹测试与生成视频对比","evidence_location":"§1、3–4，PDF pp.2–8","training_status":"有训练","annotation_notes":"生成器 checkpoint、planner 参数规模及数据总量待补充材料核实；不把 baseline 误写作所提方法 backbone。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Kizil_LAMP_Language-Assisted_Motion_Planning_for_Controllable_Video_Generation_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models","classification_review":""},{"id":"CVPR-2026-3115","venue":"CVPR","year":2026,"title":"Lighting-grounded Video Generation with Renderer-based Agent Reasoning","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Cai_Lighting-grounded_Video_Generation_with_Renderer-based_Agent_Reasoning_CVPR_2026_paper.html","doi":"","generation_model":"Wan2.2-5B + proxy encoder / adapter / LoRA","auxiliary_models":"Scene Agent + Camera Agent；具体 LLM checkpoint 在已读主文中未明确；Qwen2.5-VL 用于数据 caption","conditions":"文本；物体布局、光照、camera trajectory 经渲染形成 diffuse / glossy-GGX / rough-GGX scene proxy；支持图像／视频条件","training_datasets":"LiVERSet：LiVER-Real + LiVER-Syn；10k 训练、1k 测试，真实／合成各半","evaluation_datasets":"LiVERSet held-out 1k","evidence_location":"§3–4，PDF pp.3–6","training_status":"有训练","annotation_notes":"训练分阶段更新条件模块和 LoRA；不要把 caption VLM 自动认作 Scene Agent 的模型。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Cai_Lighting-grounded_Video_Generation_with_Renderer-based_Agent_Reasoning_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"CVPR-2026-3642","venue":"CVPR","year":2026,"title":"Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO","reasoning_role":"R2","topic":"通用生成框架 / 联合音视频","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Cheng_Video-as-Answer_Predict_and_Generate_Next_Video_Event_with_Joint-GRPO_CVPR_2026_paper.html","doi":"","generation_model":"VANS：Wan2.1-1.3B video diffusion model","auxiliary_models":"Qwen2.5-VL-3B；Gemini-2.5-Flash 用于构建 QA / CoT 数据","conditions":"输入视频 + procedural / predictive question；VLM 的 reasoning／next-event 表示连接视频回答生成","training_datasets":"VANS-Data-100K；来自 Internet 与 NEP 数据，经切片、筛选和 QA 生成；联合训练与 Joint-GRPO","evaluation_datasets":"独立 VNEP procedural / predictive 测试集；无视频／问题交叠","evidence_location":"§3、5.1，Figure 3，PDF pp.3,6","training_status":"有训练","annotation_notes":"训练集规模为数据集名所示 100K；各训练阶段拆分需见 Appendix C。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Cheng_Video-as-Answer_Predict_and_Generate_Next_Video_Event_with_Joint-GRPO_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-3739","venue":"CVPR","year":2026,"title":"Chain of Event-Centric Causal Thought for Physically Plausible Video Generation","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Wang_Chain_of_Event-Centric_Causal_Thought_for_Physically_Plausible_Video_Generation_CVPR_2026_paper.html","doi":"","generation_model":"CogVideoX-5B","auxiliary_models":"GPT-OSS-20B 做语言推理；Qwen-Image family 生成事件 keyframes","conditions":"文本物理事件 → 受公式约束的 event chain + scene graphs → 正负文本及时间对齐的 keyframe latent priors","training_datasets":"主文方法是基于现有模型的推理流程，未报告新增训练集","evaluation_datasets":"PhyGenBench、VideoPhy","evidence_location":"§3–4.1，PDF pp.4–5","training_status":"未报告新增训练","annotation_notes":"不能将 PhyGenBench / VideoPhy 当作训练数据；未从主文推断底座预训练语料。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Wang_Chain_of_Event-Centric_Causal_Thought_for_Physically_Plausible_Video_Generation_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-3787","venue":"CVPR","year":2026,"title":"TiViBench: Benchmarking Think-in-Video Reasoning for Video Generation","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Chen_TiViBench_Benchmarking_Think-in-Video_Reasoning_for_Video_Generation_CVPR_2026_paper.html","doi":"","generation_model":"被测：Wan2.2-I2V-A14B、Wan2.1-I2V-14B、HunyuanVideo-I2V、CogVideoX1.5-I2V；Veo3.1-fast、Sora2、Kling2.1","auxiliary_models":"Gemini-2.5-Pro 辅助任务 prompt 构建；VideoTPO 是推理时 prompt 优化","conditions":"任务初始图像 + 描述规则／目标的文本；可由生成结果迭代优化 prompt","training_datasets":"无新增权重训练；TiViBench 是评测集，VideoTPO 不需要额外 tuning","evaluation_datasets":"TiViBench：24 类任务、3 级难度","evidence_location":"§3–5.1、Table 1–2，PDF pp.4–6","training_status":"评测／Training-free","annotation_notes":"VideoTPO 实验应用于 Wan2.1-I2V-14B 和 HunyuanVideo-I2V。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Chen_TiViBench_Benchmarking_Think-in-Video_Reasoning_for_Video_Generation_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-3841","venue":"CVPR","year":2026,"title":"VISTA: A Test-Time Self-Improving Video Generation Agent","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Long_VISTA_A_Test-Time_Self-Improving_Video_Generation_Agent_CVPR_2026_paper.html","doi":"","generation_model":"Veo-3.0-generate-preview；Veo2 为较弱生成器实验","auxiliary_models":"Gemini-2.5-flash-preview-05-20 负责规划、比较、批评、重写","conditions":"用户文本意图 → temporal plan → prompt；生成视频的视觉／音频反馈用于下一轮","training_datasets":"无新增训练：test-time self-improvement","evaluation_datasets":"单场景 benchmark 抽取 100 prompts；内部多场景 161 prompts","evidence_location":"§3–4.1，PDF pp.3–5","training_status":"Training-free","annotation_notes":"内部 prompts 为评测而非训练 dataset。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Long_VISTA_A_Test-Time_Self-Improving_Video_Generation_Agent_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2026-3919","venue":"CVPR","year":2026,"title":"Reasoning Diffusion for Unpaired Test Time Out-of-distribution Text-Image to Video Generation","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Pan_Reasoning_Diffusion_for_Unpaired_Test_Time_Out-of-distribution_Text-Image_to_Video_CVPR_2026_paper.html","doi":"","generation_model":"ReasonDiff：Wan2.1 系列 + AlignFormer；主文所读部分未明确所用大小","auxiliary_models":"VisionNarrator；数据 caption 明确使用 Llama-3.2-11B-Vision-Instruct，推理模块的精确配置待补充材料确认","conditions":"语义未配对的文本 + 任意时刻参考图；规划逐帧叙述及图像在视频中的 anchor position","training_datasets":"WebVid；每 0.2s 取帧并由 Llama-3.2-11B-Vision-Instruct 生成 frame captions；第一阶段联合训练 base model 与 AlignFormer，第二阶段冻结 base model、仅微调 AlignFormer","evaluation_datasets":"ActivityNet 随机 500 videos 构造 unpaired text-image 测试；另用 MSR-VTT 测 paired general generation","evidence_location":"§3.4–4.2，PDF pp.5–6","training_status":"有训练","annotation_notes":"题目中的 test time 描述 OOD 输入场景，并不表示 training-free。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Pan_Reasoning_Diffusion_for_Unpaired_Test_Time_Out-of-distribution_Text-Image_to_Video_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"CVPR-2026-4022","venue":"CVPR","year":2026,"title":"CoT-Edit: Let CoT Guide Instruction Video Editing","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://openaccess.thecvf.com/content/CVPR2026/html/Liang_CoT-Edit_Let_CoT_Guide_Instruction_Video_Editing_CVPR_2026_paper.html","doi":"","generation_model":"Wan2.2-5B editor + Mask / Reverse Connectors","auxiliary_models":"Qwen-VL planner（主文明确系列，精确版本不确定）","conditions":"源视频 + 编辑指令 → CoT、逐帧 boxes、属性补充指令 → masks 与视觉语言特征条件","training_datasets":"Mask 分支：ADE20K、PhraseCut、YouTube-VOS、OVIS；Editor：AnyEdit、UltraEdit、SEED-Data-Edit、EditWorld、Señorita2M、Ditto 和自建编辑数据；联合阶段约 100k 内部有 mask 的 before/after pairs","evaluation_datasets":"视频编辑的质量、指令遵循及人评","evidence_location":"§3–4.1，PDF pp.5–6","training_status":"有训练","annotation_notes":"分支训练和联合训练数据分开记录，不能把所有数据当作一个同质训练集。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2026/papers/Liang_CoT-Edit_Let_CoT_Guide_Instruction_Video_Editing_CVPR_2026_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"ICLR-2026-1627","venue":"ICLR","year":2026,"title":"Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning","reasoning_role":"R2","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=wPEIStHxYH","doi":"","generation_model":"Cosmos-Predict2-2B，全参数 post-training","auxiliary_models":"同一模型生成 action、future state、value；无额外独立 LLM planner","conditions":"任务文本、多视角图像、proprioception；候选 action 条件用于 future-state/value 预测","training_datasets":"目标平台 robot demonstrations：LIBERO、RoboCasa、真实 ALOHA（四任务共 185 demos）；规划阶段另用本策略收集的 rollout 数据","evaluation_datasets":"LIBERO、RoboCasa、ALOHA 真实双臂任务","evidence_location":"§3–4、Appendix A.2，PDF pp.1,7,17","training_status":"有训练","annotation_notes":"这里 LIBERO / RoboCasa 既提供训练 demonstrations 也用于环境测试；不是仅凭 benchmark 名推定。","arxiv_id":"2601.16163","evidence_url":"https://arxiv.org/pdf/2601.16163","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2026-2063","venue":"ICLR","year":2026,"title":"World-In-World: World Models in a Closed-Loop World","reasoning_role":"R3","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=yDmb7xAfeb","doi":"","generation_model":"被测视频模型包括 SVD-1.5B、LTX-Video-2B、Hunyuan-13B、Wan2.1-14B、Wan2.2-5B/A14B、Cosmos-Predict2-2B、NWM、Runway-Gen4；另有图像 WM 对照","auxiliary_models":"统一 online planner 与 action API；模型配置按 zero-shot / post-trained 分开","conditions":"当前 RGB／panorama + 导航离散动作或 manipulation 的连续 7-DoF action；部分原始模型用 action-to-text 接口","training_datasets":"有 post-training 实验：Habitat-Sim 自采 action-observation 数据；RLBench 每 manipulation task 200 demonstrations；不是所有被测模型都训练","evaluation_datasets":"Active Recognition、ImageNav、Active EQA、Robotic Manipulation","evidence_location":"§2.4、Appendix C–D、Tables 6–7，PDF pp.27–29","training_status":"基准含模型训练","annotation_notes":"底座 checkpoint 和数据用途以 Table 6/7 区分；不将所有 benchmark 标为无训练。","arxiv_id":"2510.18135","evidence_url":"https://arxiv.org/pdf/2510.18135","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2026-2954","venue":"ICLR","year":2026,"title":"Dual-IPO: Dual-Iterative Preference Optimization for Text-to-Video Generation","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=mu8sO1Vw0C","doi":"","generation_model":"CogVideoX-2B / 5B 与 WanX-1.3B（按本文命名）","auxiliary_models":"VILA 13B / 40B CoT reward model；Qwen2.5 辅助构造 prompts","conditions":"文本 prompt；CoT、self-consistency 与 preference certainty 提供 reward，迭代优化生成器","training_datasets":"约 1M 组合式合成 captions；每轮生成约 100K top-bottom preference pairs；VidGen-1M 子集用于 real-data regularization；critic 用少量 CoT warm-start 与每轮约 20K pseudo-labels","evaluation_datasets":"VBench；人工偏好比较；1,000 videos 的 critic reference set","evidence_location":"§3–4.1，PDF pp.4–6；Tables 2–3，p.8","training_status":"reward model 与 generator 交替 post-training","annotation_notes":"已补读 arXiv:2502.02088 正文，替换先前摘要占位；VILA 是 critic，不能当作视频生成 backbone。","arxiv_id":"2502.02088","evidence_url":"https://arxiv.org/pdf/2502.02088","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2026-3626","venue":"ICLR","year":2026,"title":"BindWeave: Subject-Consistent Video Generation via Cross-Modal Integration","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=FP2XNyV9WL","doi":"","generation_model":"Wan 系列 DiT video foundation model；本文所读配置段落未明确精确版本／规模","auxiliary_models":"Qwen2.5-VL-7B + T5；MLP connector 连接 MLLM 与 DiT","conditions":"文本 + 单／多主体参考图；MLLM 的 subject-aware hidden states，结合 CLIP/VAE features","training_datasets":"OpenS2V-5M 过滤得到约 1M video-text pairs；两阶段 curriculum fine-tuning","evaluation_datasets":"OpenS2V benchmark","evidence_location":"§3–4，PDF pp.6–7","training_status":"有训练","annotation_notes":"OpenS2V-5M 是训练来源；OpenS2V benchmark 是评测，需分别看待。","arxiv_id":"2510.00438","evidence_url":"https://arxiv.org/pdf/2510.00438","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"ICLR-2026-4101","venue":"ICLR","year":2026,"title":"Captain Cinema: Towards Short Movie Generation","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://openreview.net/forum?id=zlNZBxQZIC","doi":"","generation_model":"FLUX.1-dev 做 interleaved keyframe generation；Seaweed-3B 做 bottom-up video generation","auxiliary_models":"Gemini-2.5 用于示例 storyline 生成，不等于全部训练 caption 模型","conditions":"详细故事文本 → 多镜头 keyframes；历史压缩视觉上下文及 text-keyframe interleaving","training_datasets":"专门整理的 cinematic interleaved text-image/video 数据；确切来源组成／数量在所读部分不确定","evaluation_datasets":"多镜头叙事视频的质量、一致性与语义对齐","evidence_location":"§3–4、Appendix A，PDF pp.3–6,16","training_status":"有训练","annotation_notes":"不是 training-free 多 agent wrapper；两个生成阶段均进行 fine-tuning。","arxiv_id":"2507.18634","evidence_url":"https://arxiv.org/pdf/2507.18634","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICLR-2026-4273","venue":"ICLR","year":2026,"title":"FilMaster: Bridging Cinematic Principles and Generative AI for Automated Film Generation","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://openreview.net/forum?id=ovSDneawKY","doi":"","generation_model":"Kling Elements，支持多张参考图；ElevenLabs 用于 voice-over","auxiliary_models":"GPT-4o（script / RAG / editing / sound design）；Gemini-2.0-Flash（audience review、foley）；Gemini-1.5-Flash 为评测器","conditions":"用户故事／参考图；检索影片片段指导 camera language，随后迭代剪辑和音轨组织","training_datasets":"未报告新增模型权重训练；440k film clips 为 RAG 参考库，46,826 audio assets 为检索音频库","evaluation_datasets":"FilmEval：20 cases，其中 10 来自 MoviePrompts","evidence_location":"§3–4、Appendix B，PDF pp.7,13","training_status":"检索／推理流程","annotation_notes":"RAG corpus 不是 training dataset；learning cinematography 在这里不可直接等同参数训练。","arxiv_id":"2506.18899","evidence_url":"https://arxiv.org/pdf/2506.18899","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2026-4471","venue":"ICLR","year":2026,"title":"AUHead: Realistic Emotional Talking Head Generation via Action Units Control","reasoning_role":"R1","topic":"人物动画 / Talking Head / HOI","url":"https://openreview.net/forum?id=dmzlAUkulz","doi":"","generation_model":"Hallo V1、MEMO + AU adapter","auxiliary_models":"Qwen-Audio-Chat 经 LoRA 训练，采用 emotion-then-AU reasoning","conditions":"audio + reference image + 预测／可编辑的 AU sequences / 2D facial representations","training_datasets":"MEAD、CREMA 的 subject-independent train splits；先训练 AU prediction，再训练视频 AU adapter","evaluation_datasets":"MEAD / CREMA 对应独立身份 test splits","evidence_location":"§3.4–4.1，PDF pp.6–7；Appendix C p.18","training_status":"有训练","annotation_notes":"避免把全部 10,000 / 7,442 clips 当作训练数量；文中先按身份划分。","arxiv_id":"2602.09534","evidence_url":"https://arxiv.org/pdf/2602.09534","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2026-4937","venue":"ICLR","year":2026,"title":"Compositional Visual Planning via Inference-Time Diffusion Scaling","reasoning_role":"R2","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=EEONns7ae4","doi":"","generation_model":"自训短程 DiT-L / DiT-XL video diffusion model + Cosmos tokenizer","auxiliary_models":"MLP inverse dynamics model 将视频计划转为低层动作","conditions":"start image + goal image；重叠视频 chunk 的边界约束与 message passing","training_datasets":"ManiSkill 派生的组合操作 demonstrations，以及真实操作示范；短程模型从长程 demos 抽短片训练，长程组合无需再训练","evaluation_datasets":"100 个组合任务（18 ID、82 OOD）及真实场景","evidence_location":"§4–5、Appendix 实现，PDF pp.7–8,22","training_status":"底座有训练／组合推理 Training-free","annotation_notes":"论文的 training-free 指组合算法，并非整套实验不需要训练数据。","arxiv_id":"2603.02646","evidence_url":"https://arxiv.org/pdf/2603.02646","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-0460","venue":"NeurIPS","year":2026,"title":"ReGen: Agentic Video World Modeling with Synergized Reasoning and Generation","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=QMvrZxy8W9","doi":"","generation_model":"causal diffusion generator；具体 checkpoint 不确定","auxiliary_models":"streaming vision-language planner；具体 checkpoint 不确定","conditions":"历史上下文 → structured action-state specification；planner 验证生成结果后决定继续、结束或重生成","training_datasets":"自建 action-grounded video dataset，超过 2M segments；具体源数据与训练拆分尚未核实","evaluation_datasets":"长程叙事连贯性／自主 rollout","evidence_location":"官方摘要；未取得可核对正文","training_status":"有训练／细节待核实","annotation_notes":"作者主页的 PDF 链接未能可靠定位对应正文；未知项不是“论文未披露”。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=QMvrZxy8W9","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2026-0902","venue":"NeurIPS","year":2026,"title":"Seeing the World through Any Eyes","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://openreview.net/forum?id=GEPHFHPezV","doi":"","generation_model":"EgoEye；底座型号不确定","auxiliary_models":"reward-guided egocentric context reasoner；具体模型不确定","conditions":"单条 exocentric 源视频；推断第一视角内容并进行跨视角 motion alignment","training_datasets":"EgoScape：19.5K 同步 exo-ego pairs + 24.2K in-the-wild egocentric videos；实际训练划分不确定","evaluation_datasets":"EgoScape 的跨视角视频生成测试","evidence_location":"官方摘要；未取得可核对正文","training_status":"有训练／细节待核实","annotation_notes":"记录的是数据集总量，不宣称全部用于训练。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=GEPHFHPezV","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2026-1384","venue":"NeurIPS","year":2026,"title":"RISE-Video: Can Video Generators Decode Implicit World Rules?","reasoning_role":"R3","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=LriqPuIS7K","doi":"","generation_model":"被测 11 模型：Hailuo2.3、Wan2.6、Sora2、Veo3.1、Kling2.6、Seedance1.5-Pro；Wan2.2-I2V-A14B/TI2V-5B、HunyuanVideo1.5-720P-I2V 及 cfg-distill、CogVideoX1.5-5B","auxiliary_models":"GPT-5 为推理／时序／物理评测 judge","conditions":"初始图像 + 含隐含规则的文本任务","training_datasets":"无新增生成模型训练；使用现成模型推理","evaluation_datasets":"RISE-Video：467 个人工标注 TI2V cases，8 类推理","evidence_location":"§3–4、模型结果表，PDF pp.5–9","training_status":"评测／无新增训练","annotation_notes":"被测模型底座训练语料不能由这篇基准推定。","arxiv_id":"2602.05986","evidence_url":"https://arxiv.org/pdf/2602.05986","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-2519","venue":"NeurIPS","year":2026,"title":"Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=XbmmVXqaHc","doi":"","generation_model":"WorldDrive：CogVideoX 初始化的 3D VAE / DiT + trajectory-aware driving WM","auxiliary_models":"共享视觉／运动 encoder 的 multi-modal planner；Future-aware Rewarder","conditions":"历史驾驶帧 + trajectory vocabulary / ego state；控制未来场景及规划候选轨迹","training_datasets":"WM：nuPlan 训练 200k iterations 后 nuScenes 100k；planner：NAVSIM navtrain 50 epochs","evaluation_datasets":"NAVSIM、NAVSIM-v2、nuScenes validation","evidence_location":"§3.4–4.3、Appendix §6.4，PDF pp.5,14","training_status":"有训练","annotation_notes":"正文概述为联合数据来源；附录进一步明确顺序训练。","arxiv_id":"2603.14948","evidence_url":"https://arxiv.org/pdf/2603.14948","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-2747","venue":"NeurIPS","year":2026,"title":"Hierarchical Denoising For Multi-Step Visual Reasoning","reasoning_role":"R2","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openreview.net/forum?id=apazc2xK3R","doi":"","generation_model":"Wan2.2-5B-TI2V + hierarchical latent tree / sparse attention","auxiliary_models":"无独立外部 LLM planner；分层 diffusion 执行 reasoning","conditions":"任务首帧；层级 latent 先规划粗粒度状态再生成细节","training_datasets":"自建 18,000 条 synthetic reasoning videos；maze、Hanoi、one-line drawing、sliding puzzle、Sokoban、water pouring 各 3,000","evaluation_datasets":"六任务的分级多步推理 benchmark，含 OOD","evidence_location":"§4.1、Appendix I，PDF pp.6,29","training_status":"有训练","annotation_notes":"全量／少数据实验使用同一任务数据体系；不是自然视频数据直接测得的零样本能力。","arxiv_id":"2607.15278","evidence_url":"https://arxiv.org/pdf/2607.15278","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-3007","venue":"NeurIPS","year":2026,"title":"OptiWorld: Optimal Control for Video World Generation under Physical Constraints","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=ElToIAx2OM","doi":"","generation_model":"视频生成 backbone 不确定；OptiWorld 是外置 optimal-control framework","auxiliary_models":"world-state extraction + manifold 上的 constrained trajectory planner；具体实现不确定","conditions":"图像／源视频、目标及物理约束 → 优化轨迹 → 视频条件","training_datasets":"摘要称控制层作用于 inference time；是否有其他组件训练及 dataset 尚未核实","evaluation_datasets":"goal-conditioned I2V、dynamics editing、counterfactual generation","evidence_location":"官方摘要；未取得可核对正文","training_status":"训练情况不确定","annotation_notes":"不能只凭 inference-time 字样把整个系统标为完全 training-free。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=ElToIAx2OM","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2026-3426","venue":"NeurIPS","year":2026,"title":"PlasticMem: Adding Temporal Reasoning to Diffusions for Consistent Long Video Generation","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://openreview.net/forum?id=xTHR2rP4s6","doi":"","generation_model":"PlasticMem diffusion renderer；底座型号不确定","auxiliary_models":"Temporal Reasoning-based Rendering 的脚本／动作规划模块；具体型号不确定","conditions":"脚本和历史视觉记忆；逐动作进行 reasoning 与独立 rendering","training_datasets":"短视频训练；数据集名称、数量和来源尚未核实","evaluation_datasets":"分钟级长程生成与 visual memory benchmark","evidence_location":"官方摘要；未取得可核对正文","training_status":"有训练／数据待核实","annotation_notes":"明确短视频训练，不推定使用任何常见 WebVid / OpenVid。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=xTHR2rP4s6","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2026-3727","venue":"NeurIPS","year":2026,"title":"Agentic Video Editing from Underspecified Requests","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://openreview.net/forum?id=pGPKhPBqR0","doi":"","generation_model":"Aurora unified video DiT；具体 backbone 不确定","auxiliary_models":"tool-augmented VLM agent；具体 checkpoint 不确定","conditions":"不完整用户请求 + 源视频；agent 补充文本、identity reference、空间目标／mask 后送入编辑器","training_datasets":"executable edit planning / reference selection 的监督数据，以及 tool use / instruction refinement preference pairs；名称及规模未核实","evaluation_datasets":"EditVerse-Bench、OpenVE-Bench、AgentEdit-Bench","evidence_location":"官方摘要；未取得可核对正文","training_status":"有训练／数据待核实","annotation_notes":"agent 的训练与向 frozen editor 迁移必须区分；不能把列出的三个 benchmark 当训练集。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=pGPKhPBqR0","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2026-3987","venue":"NeurIPS","year":2026,"title":"SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=Eez8xYYlu3","doi":"","generation_model":"支持多个 T2V/I2V：Wan2.1、Wan2.2、CogVideoX、SVI 等；闭源 Sora、Seedance1.0、Kling2.1 用于 pipeline 适配实验","auxiliary_models":"PlanAgent / CriticAgent：Qwen3-VL-8B；Gemini-3-Pro 提供 critic SFT 教师判断","conditions":"高层动作目标 + 当前／历史视频；结构化 sub-actions、memory context 与 critic feedback","training_datasets":"ActVideoGen-Dataset 用于 planner IT/DPO、video streaming long-tuning；critic 蒸馏数据 + GAIA pairwise RM；generator 另做 GRPO self-evolving","evaluation_datasets":"ActVideoGen-Bench、VBench、EgoPlan-Bench、VideoGen-RewardBench","evidence_location":"§5、Appendix B.1–B.5，PDF pp.6,14–19","training_status":"有训练","annotation_notes":"闭源模型仅接入推理流程，不能与可训练底座混为一谈；各底座精确大小见论文配置。","arxiv_id":"2603.08403","evidence_url":"https://arxiv.org/pdf/2603.08403","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-4117","venue":"NeurIPS","year":2026,"title":"Bernini: Latent Semantic Planning for Video Diffusion","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://openreview.net/forum?id=GDdIbzbPVF","doi":"","generation_model":"Wan2.2-A14B DiT renderer","auxiliary_models":"Qwen2.5-VL-7B semantic planner + T5 + connector","conditions":"文本、参考图／视频；planner 输出 ViT-space semantic plan；编辑保留 source VAE features","training_datasets":"20M 自构 video pairs；内部 T2I/T2V/I2I 与自构编辑／reference-video 数据；公开部分包括 UniREdit-100K、Pico-Banana-400K、OpenS2V-Top200K、Señorita-Controllable、SynCamVideo 等；四阶段训练","evaluation_datasets":"EditVerse、OpenVE、OpenS2V、VBench 及自建编辑评测","evidence_location":"§3–6.1、Table 1–2，PDF pp.7,13–15","training_status":"有训练","annotation_notes":"此处列关键来源，完整混合与阶段权重见 Table 1/2；内部语料未公开命名，不能据此称全部数据可复现。","arxiv_id":"2605.22344","evidence_url":"https://arxiv.org/pdf/2605.22344","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-4145","venue":"NeurIPS","year":2026,"title":"Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=86BxEBOd7Q","doi":"","generation_model":"VerseCrafter 视频后端；另一接口组合 Kimodo 人体动作生成与视频渲染","auxiliary_models":"Qwen2.5-VL-7B-Instruct，LoRA rank 128","conditions":"自然语言 + 粗人体 motion → actor/camera DSL → 连续 6-DoF trajectories；I2V 接口使用参考图及 3D Gaussian scaffold","training_datasets":"程序化合成 + CondensedMovies 实景的 text-motion-DSL 数据；读取的 arXiv 版称 34K trajectories，训练约 300K image-text samples","evaluation_datasets":"camera framing、motion alignment 及视频生成对比","evidence_location":"§3–4、Appendix A.5–A.6，PDF pp.1,6–9,16","training_status":"有训练","annotation_notes":"官方摘要称 30K，而所读 arXiv 版称 34K；保留版本差异，不把 300K samples 与 trajectory 数混为一谈。","arxiv_id":"2606.01900","evidence_url":"https://arxiv.org/pdf/2606.01900","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-4719","venue":"NeurIPS","year":2026,"title":"UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=SfVDhDwIKG","doi":"","generation_model":"Wan2.2-TI2V-5B，统一 mask-modulated DiT；沿用 text encoder / causal VAE","auxiliary_models":"共享 DiT 的 trajectory tokens，无独立 LLM planner","conditions":"历史 front-view frames、驾驶指令、ego velocity；可切换 video / action 条件与输出模式","training_datasets":"NAVSIM / OpenScene 的驾驶训练数据；共同监督 future video 与 trajectories","evaluation_datasets":"NAVSIM v1；nuScenes validation 为跨数据集测试","evidence_location":"§3、4.1、Appendix G，PDF pp.4,6,17","training_status":"有训练","annotation_notes":"主实验不依赖地图、boxes、depth、occupancy、LiDAR 监督；nuScenes 在这里不是训练集。","arxiv_id":"2607.05133","evidence_url":"https://arxiv.org/pdf/2607.05133","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-4979","venue":"NeurIPS","year":2026,"title":"Masked Visual Actions for Unified World Modeling","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=VHMcZoypoo","doi":"","generation_model":"Wan-Fun-Control-2.2-14B，LoRA rank 256","auxiliary_models":"Diffusion Policy 提出 action；MLP/learned inverse dynamics 用于动作恢复","conditions":"reference frame + masked visual action video；保留机器人像素可做正向模拟，保留目标物体轨迹可做逆向规划","training_datasets":"约 15 小时 masked examples，来自 DROID 真实视频 + RoboCasa simulation，含成功／失败；policy 另用 RoboCasa human demos / MimicGen","evaluation_datasets":"视频生成、policy evaluation、planning、动作恢复及真实跨 embodiment 测试","evidence_location":"§4.1–4.2、Appendix F，PDF pp.4–6,19","training_status":"有训练","annotation_notes":"被遮挡区域设灰色并经同一 VAE 编码；世界模型训练集与策略训练集分列。","arxiv_id":"2607.19343","evidence_url":"https://arxiv.org/pdf/2607.19343","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2026-5021","venue":"NeurIPS","year":2026,"title":"WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://openreview.net/forum?id=x1JYbqADOo","doi":"","generation_model":"被测：LTX2.3、Wan2.2-14B、UniVideo、HunyuanVideo1.5、Cosmos-Predict2.5、LongCat-Video；Sora2、Kling、Wan2.6、Seedance2.0、Veo3.1-Fast","auxiliary_models":"评测／reward 对齐比较包括 Qwen3.5-27B Thinking/Instruct、GPT-5.4","conditions":"initial visual state + action / task prompt；生成未来状态演变视频","training_datasets":"生成模型仅 inference，无新增训练；WorldRewardBench 为 reward-model evaluation 的偏好数据","evaluation_datasets":"WorldReasonBench 436 cases；WorldRewardBench 约 6K pairs / 1.4K unique videos","evidence_location":"§3、Appendix J、Licenses，PDF pp.5,31–32,40","training_status":"评测／无新增生成训练","annotation_notes":"不能把 preference benchmark 自动记成 reward-model training data。","arxiv_id":"2605.10434","evidence_url":"https://arxiv.org/pdf/2605.10434","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0039","venue":"ECCV","year":2026,"title":"How Far Are Video Models from True Multimodal Reasoning?","reasoning_role":"R3","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://eccv.ecva.net/virtual/2026/poster/5699","doi":"","generation_model":"被测：Wan2.2-14B、HunyuanVideo1.5、UniVideo、CogVideoX1.5-5B、LTX-2；Seedance2.0/1.5、Veo3.1-Fast、Wan2.7、Sora2","auxiliary_models":"AVE：Seed2.0-Pro 优化 prompt，Seed2.0-Lite judge；TextGrad / GEPA；Seed2.0-Pro 也用于辅助理解实验","conditions":"文本与图像／视频上下文；按模型可接收的参考数量和格式分组比较","training_datasets":"不训练生成模型；AVE 用少量人工标注进行 prompt 优化，非权重训练","evaluation_datasets":"CLVG-Bench：>1,000 metadata、6 类／47 子类","evidence_location":"§4.1–4.2，PDF pp.9–12","training_status":"评测／prompt 优化","annotation_notes":"文中 evaluator 的 training tasks 指优化任务组织，不能误记为 video-model training dataset。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/12354.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0106","venue":"ECCV","year":2026,"title":"UniDrive-WM: Unified Understanding, Planning and Generation World Model For Autonomous Driving","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/4081","doi":"","generation_model":"自建 VLM-centric UniDrive-WM；MoVQGAN discrete AR 或 continuous AR+diffusion decoder","auxiliary_models":"EVA02-L vision encoder、Vicuna-1.5 LLM、ORION QT-Former；LLM LoRA rank 16","conditions":"多视角图像、历史、perception queries、驾驶问答；预测轨迹作为未来图像生成条件","training_datasets":"Bench2Drive base split：950 scenes 训练／50 验证；nuScenes 官方 train/val；第二阶段加入 VQA 数据（具体源名待核实）","evaluation_datasets":"Bench2Drive 220 closed-loop routes；nuScenes","evidence_location":"§3–4.3，PDF pp.7–10","training_status":"有训练","annotation_notes":"以 future-frame prediction 为核心，原池按 world-model 范围收录；不等同开放域长视频生成。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4407.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ECCV-2026-0260","venue":"ECCV","year":2026,"title":"What if? Emulative Simulation with World Models for Situated Reasoning","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/4721","doi":"","generation_model":"HunyuanVideo、CogVideoX、Wan 系列；细节实验包含 Wan2.1 LoRA / Wan2.2 prompt extension","auxiliary_models":"Qwen3-VL-32B、LLaVA-OneVision-1.5-8B 负责 situated reasoning","conditions":"有限起始观察 + target situation / what-if question；相机轨迹或其动作文字扩展控制 imagined video","training_datasets":"WanderDream-Gen 的两类训练子集，主要 LoRA 8 epochs；源自 HM3D、ScanNet++ 的路径视频与真实捕获；总数据 15.8K videos，训练数量以 split 为准","evaluation_datasets":"WanderDream-Gen；WanderDream-QA 158K QA 的评测划分","evidence_location":"§4–5.1，PDF pp.9–10,13","training_status":"有训练","annotation_notes":"QA 用于评估 reasoning；不把 158K QA 全部写作 video model 的训练数据。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/7191.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0267","venue":"ECCV","year":2026,"title":"Nexus-Vid: Efficient Frequency Bridging with Homogeneous Latent Space for Video Unified Models","reasoning_role":"R1","topic":"效率 / 架构 / 蒸馏 / 实时","url":"https://eccv.ecva.net/virtual/2026/poster/5094","doi":"","generation_model":"Omni-Video 已对齐的 Wan2.1-T2V-1.3B small generator；Wan2.1-T2V-14B large generator 在推理时接续","auxiliary_models":"Omni-Video understanding block；精确 checkpoint 在已读配置段未明确；Qwen3-VL-30B-A3B-Instruct 是 evaluator","conditions":"文本／多模态理解产生 semantic control；在共享 latent 中进行频率与时间上的逐步生成器交接","training_datasets":"复用 unified small-generator 训练／对齐框架；具体训练 corpus 未从所读主文核实","evaluation_datasets":"VBench、VR-Bench（216 cases，8 dimensions）","evidence_location":"§3–4，PDF pp.3–9","training_status":"有训练／数据待核实","annotation_notes":"会议目录题名 Nexus-Vid，获取的官方 PDF 题名 Lumos-Nexus；保留该版本差异。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/8809.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"ECCV-2026-0278","venue":"ECCV","year":2026,"title":"LangDriveCTRL: Natural Language Controllable Driving Scene Editing with Multi-modal Agents","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://eccv.ecva.net/virtual/2026/poster/4542","doi":"","generation_model":"CogVideoX，初始化 TrajectoryCrafter checkpoint，再训练 video harmonization","auxiliary_models":"GPT-4 LLM + GPT-4o VLM 的 orchestration / grounding / behavior / reviewer agents","conditions":"源驾驶视频 + 自然语言编辑；3D scene graph、对象 mesh、轨迹与渲染视频作条件","training_datasets":"训练使用 mesh 替换 3DGS 渲染对和 cycle-reconstruction 生成的 noisy-rendering pairs；实际源训练数据名称尚未核实","evaluation_datasets":"Waymo Open Dataset 筛选 30 场景及 GPT-4 生成后人工筛选的编辑指令","evidence_location":"§3、4.2，PDF pp.9–11","training_status":"有训练","annotation_notes":"Waymo 30 scenes 明确用于实验评测；不能未经说明就当作完整训练集。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/6406.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ECCV-2026-0321","venue":"ECCV","year":2026,"title":"DIVER: Disentangling Camera–Object and Active–Passive Motion for Video Generation","reasoning_role":"R2","topic":"条件控制 / 定制 / 几何一致","url":"https://eccv.ecva.net/virtual/2026/poster/5541","doi":"","generation_model":"Wan2.1-14B；更新 camera/trajectory encoders 与 self-attention","auxiliary_models":"无独立 LLM planner；学习 active/passive motion 的 action–consequence 依赖","conditions":"参考图像、文本、canonical static-view active object trajectories + 独立 camera trajectory；可对被动后果推断或反推主动运动","training_datasets":"Panda-70M + Wild-SDG1M 选出 76K static-view videos；生成 43K paired dynamic-view videos；另 3.4K SyncMaster synthetic interaction videos","evaluation_datasets":"DynPose-100K、Cooking、WISA","evidence_location":"§3.3、4.1，PDF pp.9,12–13","training_status":"有训练","annotation_notes":"目录题名 DIVER，官方摘要与 PDF 方法名 MoRight；统计记录保留目录 ID。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/11272.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0382","venue":"ECCV","year":2026,"title":"InteractiveAvatar: Real-Time Streaming Video Generation for Consistent and Intent-Aware Avatars","reasoning_role":"R1","topic":"人物动画 / Talking Head / HOI","url":"https://eccv.ecva.net/virtual/2026/poster/3402","doi":"","generation_model":"Wan2.2-5B，四阶段 audio-image-to-video / autoregressive distillation；LSVM","auxiliary_models":"Reasoning-Reaction Module；具体外部模型 checkpoint 未从主文核实","conditions":"reference portrait、流式音频、用户意图／交互状态、长短视觉记忆；可切换生成 cache","training_datasets":"约 3M 清洗 clips：HDTF、VFHQ、VoxCeleb2、CelebV-Text、AVSpeech；OpenHumanVid 影视数据；自有长对话／丰富动作数据","evaluation_datasets":"长时 avatar consistency、intent-aware interaction 与实时性","evidence_location":"§3–4.2，PDF pp.6,10–12","training_status":"有训练","annotation_notes":"3M 为清洗后合并总量，各来源比例未在所读段落拆开。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1287.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0574","venue":"ECCV","year":2026,"title":"PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://eccv.ecva.net/virtual/2026/poster/4063","doi":"","generation_model":"Wan2.1-14B + LoRA-Switch Reference；groupwise DPO","auxiliary_models":"VideoCon-Physics 为训练 reward；CoT VLM 支持 PhyAugPipe 数据构建，具体构建模型待核实","conditions":"文本 prompt；真实 winning video 与多个生成候选、物理 reward 作为训练信号","training_datasets":"自建 PhyVidGen-135K；包含物理增强数据构建后的真实视频／文本与生成偏好组","evaluation_datasets":"PhyGenBench、VideoPhy2","evidence_location":"§3–4、实验配置，PDF pp.3,6,9–11","training_status":"有训练","annotation_notes":"VideoPhy2-AutoRater 是评测器，与训练 reward VideoCon-Physics 区分。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4366.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"ECCV-2026-0677","venue":"ECCV","year":2026,"title":"Benchmarking Scientific Understanding and Reasoning for Video Generation using VideoScience-Bench","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://eccv.ecva.net/virtual/2026/poster/5536","doi":"","generation_model":"被测：Sora2、Veo3、Seedance1.0-Pro、Kling-v2.5-Turbo-Pro、Hailuo2.3、Ray2、Wan2.5-T2V-Preview","auxiliary_models":"VideoScience-Judge：VLM + CV 工具；比较 GPT-4o / GPT-5-pro / Claude4.5-Sonnet / Gemini3-Pro","conditions":"科学实验文本描述；部分案例附初始实验图像","training_datasets":"无新增视频模型训练；专家构建测试任务与评分规则","evaluation_datasets":"VideoScience-Bench：160 cases，合计 200 T2V/I2V evaluation items","evidence_location":"§3–4.1，PDF pp.6–10","training_status":"评测／无新增训练","annotation_notes":"参考实验视频用于构造评测题与 ground truth，不是本论文训练语料。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/11226.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-0955","venue":"ECCV","year":2026,"title":"VOID: Video Object and Interaction Deletion","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://eccv.ecva.net/virtual/2026/poster/3342","doi":"","generation_model":"CogVideoX DiT，初始化 Generative Omnimatte 权重后微调","auxiliary_models":"Gemini-3-Pro 判定删除对象的影响区域；GPT-5.2、Qwen3.5-VL 为附录变体","conditions":"源视频 + 要删除的对象；VLM 生成 quadmask；第二遍利用 flow-warped noise 保持结构","training_datasets":"Kubric + HUMOTO 构建成对 counterfactual object-removal 视频；含删除后 collision / support 等后果变化","evaluation_datasets":"真实视频；40 个 synthetic counterfactual 测试视频，与训练独立","evidence_location":"§3.3–3.6、4.3–4.4，PDF pp.6–10,13–14","training_status":"有训练","annotation_notes":"消融中 1,200 samples 是控制规模的子集，不当作完整训练集总量。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1029.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-1141","venue":"ECCV","year":2026,"title":"Making Avatars Interact: Towards Text-Driven Human-Object Interaction for Controllable Talking Avatars","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/4099","doi":"","generation_model":"InteractAvatar：PIM motion stream + AIM audio-video stream；精确 pretrained backbone 未从主文核实","auxiliary_models":"Wav2Vec audio encoder；DWPose / DINO 用于 pose/object 标注","conditions":"文本、音频、reference image；环境对象和骨架 motion；可联合生成或外部 motion 驱动","training_datasets":"人体／物体互动视频经检测生成 motion supervision；joint-generation : external-driving 训练比例 4:1；原始训练 corpus 名称待补充材料核实","evaluation_datasets":"GroundInter：400 张 Jimeng4.0 构建的 interaction images","evidence_location":"§3–4，PDF pp.7–9","training_status":"有训练／数据源待核实","annotation_notes":"GroundInter 是新建评测，不能填作训练集。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4488.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-1189","venue":"ECCV","year":2026,"title":"Better Call CineCrew: Consistent Ultra-Long Narrative-to-Film Generation","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://eccv.ecva.net/virtual/2026/poster/3661","doi":"","generation_model":"off-the-shelf keyframe / I2V generators；支持异构后端，主文所读实验未明确统一 checkpoint","auxiliary_models":"多 agent FilmDSL planning / production / review；Gemini-3-Flash 是 evaluator","conditions":"叙事脚本 → FilmDSL；character / set assets、keyframes、continuity constraints、persistent memory","training_datasets":"无新增生成器训练；通过 structured QA 与 prompt / plan 修订执行修复","evaluation_datasets":"CineCrewEval、MovieBench 抽样的人评叙事","evidence_location":"§3.4–4，PDF pp.9–12","training_status":"Training-free orchestration","annotation_notes":"不能把 evaluator 模型或 baseline 的生成器当作系统各 agent 的 checkpoint。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/2445.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"ECCV-2026-1203","venue":"ECCV","year":2026,"title":"DriveVA: Video Action Models are Zero-Shot Drivers","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/3252","doi":"","generation_model":"Wan2.2-TI2V-5B，统一 video-action DiT 与 causal VAE","auxiliary_models":"无独立外部 LLM planner","conditions":"4 历史帧、驾驶文本与 ego context；联合预测 8 future frames / trajectories，可做 video continuation","training_datasets":"NAVSIM 驾驶训练；基于现成 video backbone full fine-tuning","evaluation_datasets":"NAVSIM；nuScenes / Bench2Drive 为不做目标域微调的 zero-shot transfer","evidence_location":"§3、4.2、Conclusion，PDF pp.8,10,15","training_status":"有训练","annotation_notes":"标题 zero-shot 指跨域迁移，不是模型无需训练。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/611.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-1501","venue":"ECCV","year":2026,"title":"Target-Bench: Can Video World Models Achieve Mapless Path Planning with Semantic Targets?","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://eccv.ecva.net/virtual/2026/poster/5197","doi":"","generation_model":"被测 Sora2、Veo3.1/fast、Wan2.5-I2V-preview、Wan2.2-plus/flash、Wan2.1-plus/turbo；另微调 Wan2.2-TI2V-5B","auxiliary_models":"VGGT、SpaTracker、ViPE 从生成视频恢复相机路径","conditions":"首帧 + 语义目标文本；要求生成走向目标的第一视角视频","training_datasets":"Target dataset 的 train split，121-frame clips；比较普通 LoRA 与 shifting-frame 四倍增广","evaluation_datasets":"Target-Bench：450 robot-collected scenarios、47 semantic categories 的测试划分","evidence_location":"§3–4.3，PDF pp.4,7,10–13","training_status":"基准含模型训练","annotation_notes":"450 是全数据场景数，不宣称全部用于 LoRA 训练。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9345.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-1512","venue":"ECCV","year":2026,"title":"Wan-R1: Verifiable-Reinforcement Learning for Generalizable Video Reasoning","reasoning_role":"R2","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://eccv.ecva.net/virtual/2026/poster/5896","doi":"","generation_model":"Wan2.2-TI2V-5B；以 maze demonstrations 微调的 Wan-SFT 为 RL 起点","auxiliary_models":"GRPO / 可验证任务奖励；机器人任务另用 reference embedding reward","conditions":"初始场景图像 + 任务文本","training_datasets":"VR-Bench 的迷宫示范用于 SFT / RL；机器人路径任务使用 Target 数据与参考轨迹；具体 split 数量未核实","evaluation_datasets":"VR-Bench：regular / irregular / 3D maze、trapfield、Sokoban；Target-Bench","evidence_location":"§4.1–4.3，PDF pp.8–9；training appendix","training_status":"SFT + RL","annotation_notes":"不把 VR-Bench 和 Target-Bench 的全部测试样本视为训练集。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/13964.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ECCV-2026-1597","venue":"ECCV","year":2026,"title":"DeforM: Reasoning-Guided Physics-Aware Video Generation via Spatial-Temporal Masking","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://eccv.ecva.net/virtual/2026/poster/3443","doi":"","generation_model":"Wan2.2-TI2V-5B","auxiliary_models":"Qwen2.5-VL-7B 物理推理；LLMDet 检测","conditions":"图像 + 文本；内部转成 physics-critical region 的时空 mask","training_datasets":"DeforM-Injection：WISA-80K 中筛选约 6K deformation clips，LoRA rank 256；DeforM-Free 不新增训练","evaluation_datasets":"形变任务，包括 melting / squeezing / fracturing；测试 split 名称未核实","evidence_location":"§4.1–4.2，PDF p.10","training_status":"Training-free 与 LoRA 两个版本","annotation_notes":"两种方法的训练需求不同，不能统一标成 training-free。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/1472.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-1602","venue":"ECCV","year":2026,"title":"EvoWorld: A World-Model-Centric Framework for Continuous Self-Evolution of Modular Embodied Skills","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/5417","doi":"","generation_model":"action-conditioned latent world model；具体生成 backbone 未核实","auxiliary_models":"Skill Router、VLM evaluator 与模块化 VLA experts；实验包含 GR00T-N1.6、PI-0.5、PI-0","conditions":"任务指令、初始场景、VLA 输出的动作序列","training_datasets":"共享 offline experience repository 与验证后的 imagined trajectories；文中以 DROID 为场景 grounding 示例；精确预训练数据配方与 split 未核实","evaluation_datasets":"5 类 manipulation skill；Franka 真机实验","evidence_location":"§3.3–4.1，PDF pp.6–10；Table 1、real-robot experiments","training_status":"模块化迭代训练","annotation_notes":"所获主文把具体预训练配置放到 Appendix，但该 PDF 未附相应实验附录；Related Work 的 OXE / RH20T 不计作已确认训练集。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/10502.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；training_datasets","classification_review":""},{"id":"ECCV-2026-1910","venue":"ECCV","year":2026,"title":"From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://eccv.ecva.net/virtual/2026/poster/4366","doi":"","generation_model":"Vid-PRE 接入 Wan2.2、Wan2.5、Wan2.6-Flash；精确 Wan2.2 checkpoint 未核实","auxiliary_models":"Qwen3-VL-8B prompt reasoner；Gemini-3.0-Pro 生成训练标注并评测","conditions":"初始图像 + 用户任务文本 → CoT 与重写后的生成 prompt","training_datasets":"自建约 20K 高质量样本做 SFT；1.5K 高难样本做 GRPO，使用 text-based reward；训练的是 reasoner，不是视频生成器","evaluation_datasets":"VWG-Bench，9 dimensions / 38 tasks","evidence_location":"§3–4，PDF pp.5–6,10","training_status":"Reasoner SFT + RL；生成器冻结","annotation_notes":"数据来源含开源 GUI / embodied tasks、程序生成题目与合成图像；不把整个 VWG-Bench 自动等同训练集。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/5598.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"ECCV-2026-1966","venue":"ECCV","year":2026,"title":"Evaluating Reasoning Coherence in Video Generative Models with Text and Visual Hints","reasoning_role":"R3","topic":"评测 / 数据 / 分析 / 安全","url":"https://eccv.ecva.net/virtual/2026/poster/5161","doi":"","generation_model":"被测 Veo-3.1、Veo-3.1-fast、Sora-2、Seedance-1.0-pro/fast、Kling-v2.1、Cosmos-Predict2-14B","auxiliary_models":"Gemini-2.5-Flash 作为 VLM judge","conditions":"初始图像 + 任务；no-hint / text-hint / visual-hint 三种设置","training_datasets":"无新增训练；使用现成生成模型的默认配置","evaluation_datasets":"MME-CoF-Pro：303 samples、16 categories","evidence_location":"§4.1，PDF p.9","training_status":"仅评测","annotation_notes":"这里的 model 是被测模型，不是作者训练的新 backbone。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9137.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2026-2217","venue":"ECCV","year":2026,"title":"LMGenDrive: Bridging Multimodal Understanding and Generative World Modeling for End-to-End Driving","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://eccv.ecva.net/virtual/2026/poster/5240","doi":"","generation_model":"Stable Diffusion 1.5 + AnimateDiff；ResNet / BEV 视觉编码","auxiliary_models":"Vicuna-7B 联合预测 actions 与 world features","conditions":"三路摄像头历史图像 + 自然语言驾驶指令；视频生成进一步受 LLM world queries 控制","training_datasets":"视觉编码器：CARLA 采集的 3M expert frames；随后联合单步训练，再冻结生成器做多步 LLM 训练；后两阶段独立数据集名称 / 规模未核实","evaluation_datasets":"CARLA 0.9.10.1 的 LangAuto / LangAuto-Short / LangAuto-Tiny","evidence_location":"§3.5–4.1，PDF pp.9–11","training_status":"三阶段训练","annotation_notes":"LangAuto 在所读段落中是评测 benchmark，不直接填成全部训练数据。","arxiv_id":"未核实","evidence_url":"https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/9549.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICML-2026-1668","venue":"ICML","year":2026,"title":"Learning Latent Action World Models in the Wild","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=HDf5semiaB","doi":"","generation_model":"冻结 V-JEPA 2-L encoder；ViT-L world model 与 ViT-L pixel decoder","auxiliary_models":"jointly trained inverse dynamics model；真实动作到 latent action 的 lightweight controller；CEM planning","conditions":"历史视频表征 + 128-D continuous latent actions；下游 controller 接真实机器人 / 导航动作","training_datasets":"YoutubeTemporal-1B 训练 latent action world model；DROID、RECON 训练下游 controller","evaluation_datasets":"DROID manipulation / RECON navigation；latent action transfer、rollout 与 planning","evidence_location":"§4、§8–9；Appendix A，PDF pp.9,12,17","training_status":"预训练 + controller 训练","annotation_notes":"世界模型主要在 representation space 预测，另训练 pixel decoder 可视化；不是直接套用视频 diffusion checkpoint。","arxiv_id":"2601.05230","evidence_url":"https://arxiv.org/pdf/2601.05230","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2026-1741","venue":"ICML","year":2026,"title":"DreamDojo: A Real-Time Robot World Model from Large-Scale Human Videos","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=FuvU7PTyED","doi":"","generation_model":"Cosmos-Predict2.5，DreamDojo 2B / 14B；700M latent action Transformer","auxiliary_models":"action MLP；real-time distillation","conditions":"历史观察 + latent actions；post-training 后使用目标机器人真实动作；text condition 固定为空","training_datasets":"世界模型：In-lab、EgoDex、DreamDojo-HV，采样比 1:2:10；latent action model 还使用 Unitree G1、Fourier GR-1、AgiBot、YAM 内部机器人数据；目标机器人数据用于 post-training","evaluation_datasets":"机器人 action-controlled rollout、In-lab Eval、EgoDex Eval 等","evidence_location":"§3.3.3、§4.1，PDF pp.7,9–11","training_status":"预训练 + 机器人 post-training + 蒸馏","annotation_notes":"会议标题是 Real-Time Robot World Model，所读 arXiv 正文标题用 Generalist Robot World Model；数据和配置按所链接版本记录。","arxiv_id":"2602.06949","evidence_url":"https://arxiv.org/pdf/2602.06949","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2026-1784","venue":"ICML","year":2026,"title":"Bridging Your Imagination with Audio-Video Generation via a Unified Director","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://openreview.net/forum?id=FHMsceUzfS","doi":"","generation_model":"UniMAGE unified director，Mixture-of-Transformers；底层 audio-video generator checkpoint 未核实","auxiliary_models":"统一 text / image generation 的 director 自身负责脚本与 keyframe 推理","conditions":"用户文本 → interleaved script / keyframes → audio-video generation","training_datasets":"摘要明确 interleaved text-image data 与 first-interleaving-then-disentangling training；具体 dataset 名称、规模未核实","evaluation_datasets":"脚本逻辑与 keyframe consistency 实验；数据集未核实","evidence_location":"官方录用摘要；未获得可读正文","training_status":"有训练，细节未核实","annotation_notes":"不依据摘要猜测其使用 Wan、Sora 或具体图像 backbone。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=FHMsceUzfS","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；training_datasets","classification_review":""},{"id":"ICML-2026-1911","venue":"ICML","year":2026,"title":"DiLA: Disentangled Latent Action World Models","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=BRBHruBDkb","doi":"","generation_model":"DiLA：DINOv2-base-with-registers encoder、RAE ViT-XL decoder 冻结；约 123M 可训练模块","auxiliary_models":"IDM / FDM、content memory；下游 action MLP 与 MPPI planning","conditions":"初始 / 历史帧 + disentangled latent actions；规划阶段映射真实动作","training_datasets":"Something-Something-v2、RT-1、RECON、LoopNav；下游 Robosuite 5,000 trajectories、RoboDesk 35,000 perturbed scripted trajectories 做 action adaptation / fine-tuning","evaluation_datasets":"SSv2 / RT-1 generation 与 transfer；RoboDesk / Robosuite MPC","evidence_location":"§4，PDF pp.5–8；Appendix A / D，pp.13,16","training_status":"训练 + action adaptation","annotation_notes":"冻结的 encoder / decoder 预训练来源与本文新训练的数据分开；未追溯其全部原始预训练语料。","arxiv_id":"2605.15725","evidence_url":"https://arxiv.org/pdf/2605.15725","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2026-3051","venue":"ICML","year":2026,"title":"MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=CKpqRFTRfc","doi":"","generation_model":"Wan2.2-TI2V-5B DiT；RGB / depth 双分支与多视角融合","auxiliary_models":"预训练 TCN-VAE action encoder / decoder；test-time latent action optimization；residual inverse dynamics","conditions":"单视角 RGB-D 首帧 + 文本指令；训练时另有动作轨迹 latent、masked source video；输出多视角 RGB-D","training_datasets":"分别在 RLBench（超过 8,000 trajectories）、RoboTwin2（超过 10,000 trajectories）与自采真实机器人数据上训练；真实数据规模未核实","evaluation_datasets":"synthetic / real 4D scene generation；两套仿真与一套真机 manipulation","evidence_location":"§4.3–5.1，PDF pp.5–6；Appendix A，p.11","training_status":"SFT + action VAE 训练；推理时优化","annotation_notes":"每个 dataset 单独训练，不是将全部数据混合成一个共同 checkpoint。","arxiv_id":"未核实","evidence_url":"https://mercerai.github.io/MVISTA-4D/icml2026.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICML-2026-4190","venue":"ICML","year":2026,"title":"Co-Evolving Latent Action World Models","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=iec4FIZyrx","doi":"","generation_model":"主要使用 OpenSora；另验证 Wan2.1-T2V-1.3B","auxiliary_models":"jointly trained inverse dynamics / vector-quantized latent action model；下游 2-layer action adapter","conditions":"初始视觉观察 + latent actions，经 AdaLN 注入；下游适配真实动作","training_datasets":"混合约 30% OXE、20% AgiBot、50% human video；后者列出 SSv2、RH20T、Ego4D、EgoPAT3D、EGTEA Gaze+、HOI4D、EPIC-KITCHENS、HO-Cap、HoloAssist；LIBERO 下游训练 split 用于 probing / adaptation","evaluation_datasets":"各数据 valid split；LIBERO 与 action-conditioned simulation / control","evidence_location":"§3.3；Appendix A–C，PDF pp.13–15；D.4，p.19","training_status":"warm-up + joint training + 下游适配","annotation_notes":"OpenSora 精确版本未核实；Wan2.1 是额外 backbone 实验。","arxiv_id":"2510.26433","evidence_url":"https://arxiv.org/pdf/2510.26433","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2026-5339","venue":"ICML","year":2026,"title":"ReaForest: Fostering Generative Video Reasoning for Spatial Planning","reasoning_role":"R2","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=Ktvb8rqFV7","doi":"","generation_model":"Video Generation Model backbone 未核实","auxiliary_models":"ReaTTS 多分支搜索、reflective recovery 与 hierarchical constraint verifier；具体辅助模型未核实","conditions":"spatial planning 任务输入；精确图像 / 文本接口未核实","training_datasets":"ReaGen-27k，摘要明确用于激活空间规划推理能力；训练配方未核实","evaluation_datasets":"空间规划；摘要列 Gemini-2.5-Pro、Sora-2 为比较对象","evidence_location":"官方录用摘要；作者主页指向 OpenReview，正文访问未成功","training_status":"有训练 + test-time scaling，细节未核实","annotation_notes":"Gemini / Sora 是摘要中的比较对象，不能当作本文采用的 backbone。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=Ktvb8rqFV7","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；conditions；training_datasets","classification_review":""},{"id":"ICML-2026-6035","venue":"ICML","year":2026,"title":"MotiMotion: Motion-Controlled Video Generation with Visual Reasoning","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=ndyuCnJxks","doi":"","generation_model":"Wan2.2-I2V-A14B","auxiliary_models":"Gemini-3.1-Pro motion reasoning；CoTracker3 生成训练 point trajectories","conditions":"初始图像、文本、用户稀疏轨迹；reasoner 修正主运动并补充 secondary motion；confidence-aware point maps","training_datasets":"OpenVid：先 5K steps 学 motion control，再 3K steps 对 50% 样本扰动轨迹，学习 confidence-aware control","evaluation_datasets":"MotiBench：collision、constraint change、tool mechanisms、flow、common objects","evidence_location":"§3.1–4.1，PDF pp.4–6；Appendix A，p.13","training_status":"视频控制模块 fine-tuning；reasoner training-free","annotation_notes":"training-free 仅适用于 VLM reasoner，不能据此把整篇标为不训练。","arxiv_id":"2605.22818","evidence_url":"https://arxiv.org/pdf/2605.22818","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ECCV-2024-1178","venue":"ECCV","year":2024,"title":"VideoStudio: Generating Consistent-Content and Multi-Scene Videos","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://vidstudio.github.io/","doi":"","generation_model":"VideoStudio-Img 基于 SD2.1；VideoStudio-Vid 基于 SDXL 加 temporal attention / convolution","auxiliary_models":"ChatGLM3-6B 编写脚本；U2-Net 分离前景背景；CLIP 编码参考图像","conditions":"文本 → 分镜脚本、实体描述、相机运动；foreground / background reference images 与 scene-reference image","training_datasets":"VideoStudio-Img 新 cross-attention：LAION-2B；VideoStudio-Vid：WebVid-10M、HD-VG-130M","evaluation_datasets":"MSR-VTT、WebVid-10M validation、Coref-SV / Pororo-SV 衍生多镜头测试","evidence_location":"§3、§4.1–4.4，PDF pp.5–10","training_status":"图像与视频模块训练；LLM 推理","annotation_notes":"GPT-4 是文中讨论的可替代脚本模型，主要实现使用 ChatGLM3-6B。","arxiv_id":"未核实","evidence_url":"https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/07783.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2024-1470","venue":"ICLR","year":2024,"title":"LLM-grounded Video Diffusion Models","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://openreview.net/forum?id=exKHibougU","doi":"","generation_model":"ModelScope text-to-video diffusion","auxiliary_models":"GPT-3.5 / GPT-4 dynamic scene layout planner","conditions":"文本 → 各帧 object boxes / dynamic scene layouts → attention guidance","training_datasets":"无新增权重训练；WebVid-2M 用于 retrieval baseline / in-context examples，不是本文模型训练集","evaluation_datasets":"自建 500 prompts 的五类时空 benchmark；UCF-101、MSR-VTT","evidence_location":"§4–5，PDF pp.6–8；Appendix A","training_status":"Training-free","annotation_notes":"ModelScope 原始预训练语料未作为本文新增训练数据。","arxiv_id":"2309.17444","evidence_url":"https://arxiv.org/pdf/2309.17444","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2024-1627","venue":"ICLR","year":2024,"title":"Video Language Planning","reasoning_role":"R2","topic":"通用生成框架 / 联合音视频","url":"https://openreview.net/forum?id=9pKtcJcMP3","doi":"","generation_model":"UniPi 风格多尺度 pixel-space video diffusion：24×40 → 48×80 → 192×320","auxiliary_models":"PaLM-E 12B policy / value；LAVA 风格 goal-conditioned policy","conditions":"当前图像 + 长程任务文本；tree search 中短程语言动作作为视频条件","training_datasets":"Language Table；RT-1；约 1,200 双臂 kitchen stacking teleoperation demonstrations；generalization 实验另混合自采 7DoF、Bridge、RT-2、Ego4D、EPIC-KITCHEN、LAION-400M","evaluation_datasets":"Language Table、7DoF mobile manipulation、14DoF bimanual manipulation","evidence_location":"Appendix A.3–A.5，PDF pp.14–16","training_status":"按 domain 训练 video / VLM / policy","annotation_notes":"generalization 数据混合不应推广成所有实验都用的统一训练集。","arxiv_id":"2310.10625","evidence_url":"https://arxiv.org/pdf/2310.10625","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2025-2467","venue":"ICLR","year":2025,"title":"GEVRM: Goal-Expressive Video Generation Model For Robust Visual Manipulation","reasoning_role":"R2","topic":"通用生成框架 / 联合音视频","url":"https://openreview.net/forum?id=hPWWXpCaJ7","doi":"","generation_model":"预训练 text-guided video DiT（引用 Zheng et al., 2024）；T5 encoder；精确 checkpoint 未核实","auxiliary_models":"ResNet34 visual encoders；goal-guided diffusion policy 与 internal-model feedback","conditions":"当前视觉观察 + 语言目标；预测未来 visual goals，再以目标与当前状态控制动作","training_datasets":"CALVIN、Bridge 的 training split；真机自采 demonstration 数据","evaluation_datasets":"CALVIN、Bridge 测试划分与外部扰动；真机 manipulation","evidence_location":"§4–5.1，PDF pp.5–7；Appendix real-world，p.16","training_status":"video planner / VAE / goal policy 训练","annotation_notes":"仅由参考文献推测 OpenSora 的精确版本不够可靠，保留 checkpoint 未核实。","arxiv_id":"2502.09268","evidence_url":"https://arxiv.org/pdf/2502.09268","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"ICLR-2025-2989","venue":"ICLR","year":2025,"title":"COMBO: Compositional World Models for Embodied Multi-Agent Cooperation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=YXRyYkb1im","doi":"","generation_model":"自训 compositional video diffusion U-Net + inpainting / super-resolution；128² → 336²","auxiliary_models":"LLaVA-v1.5-7B planning modules；T5-XXL text encoder","conditions":"局部 egocentric observations → overall world state；多个 agent 的文本动作组合","training_datasets":"脚本 rollout 采集 TDW-Game 107K videos、TDW-Cook 50K videos；另采 4K training environments 的 planning supervision 微调 VLM","evaluation_datasets":"TDW 合作任务，2–4 agents","evidence_location":"§3；Appendix B.1–B.2，PDF pp.19–22","training_status":"两阶段 compositional diffusion + VLM fine-tuning","annotation_notes":"生成器数据与 planner 监督分开记录。","arxiv_id":"2404.10775","evidence_url":"https://arxiv.org/pdf/2404.10775","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2025-3077","venue":"ICLR","year":2025,"title":"FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=B2N0nCVC91","doi":"","generation_model":"flow-conditioned latent video diffusion Transformer，使用 SDXL VAE；另有 flow diffusion proposal model","auxiliary_models":"LIV vision-language value model；diffusion-style low-level policy；Meta Llama 3.1 8B 提供 policy 文本 tokens","conditions":"图像历史 + 语言目标；dense flow plans 条件化视频生成","training_datasets":"LIBERO-LONG（50×10 videos）、FMB（1K single-object / 100 multi-object videos）、自采 folding / unfolding（各 40 videos）；另有 Bridge-V2 模块实验、LIBERO-90 预训练变体；policy 使用少量 action-labeled demos","evaluation_datasets":"LIBERO-LONG、FMB、cloth tasks、Bridge-V2；真机操作","evidence_location":"§5，PDF pp.7–10；Appendix A.1–A.3，p.16；B.2，p.17","training_status":"flow / dynamics / value / policy 训练","annotation_notes":"SDXL 是 VAE 来源，不能写成整个视频生成器是现成 SDXL checkpoint。","arxiv_id":"2412.08261","evidence_url":"https://arxiv.org/pdf/2412.08261","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICLR-2025-3138","venue":"ICLR","year":2025,"title":"VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://openreview.net/forum?id=Kb9PnkWYNT","doi":"","generation_model":"以 VideoCrafter2 为主要展示的 pretrained T2V diffusion；适用于 temporal attention 架构","auxiliary_models":"ChatGPT / LLM 生成 Video Synopsis Prompting；精确 API snapshot 未核实","conditions":"单条动态过程文本 → 多阶段描述及插值文本 embeddings；推理时 temporal attention regularization","training_datasets":"无新增模型训练","evaluation_datasets":"动态过程 prompts、长视频与 temporal dynamics 比较；具体 benchmark 名称未核实","evidence_location":"§3–4，PDF pp.4–8","training_status":"Training-free inference","annotation_notes":"LaVie / AnimateDiff 等有比较实验，未一概当作该方法主 backbone。","arxiv_id":"2403.13501","evidence_url":"https://arxiv.org/pdf/2403.13501","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"ICML-2024-1859","venue":"ICML","year":2024,"title":"RoboDreamer: Learning Compositional World Models for Robot Imagination","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"","doi":"","generation_model":"自训 compositional video diffusion U-Net；64² base + 128² / 256² upsamplers","auxiliary_models":"冻结 T5-XXL；Stable Diffusion VQ-VAE image encoder；ResNet18 inverse dynamics","conditions":"当前观察、分解后的文本 primitives；可附 goal image / sketch 等多模态目标","training_datasets":"RT-1 robot demonstrations，约 70K；按组合条件训练 video world model；inverse dynamics 用相邻帧与机器人状态 / 动作监督","evaluation_datasets":"RT-1 / robot manipulation 的 compositional generalization","evidence_location":"§3–4，PDF pp.5–7；Appendix A，p.11","training_status":"world model + inverse dynamics 训练","annotation_notes":"Stable Diffusion 在此用于图像条件编码，不能将其称为整套生成器 checkpoint。","arxiv_id":"未核实","evidence_url":"https://raw.githubusercontent.com/mlresearch/v235/main/assets/zhou24f/zhou24f.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2024-2015","venue":"ICML","year":2024,"title":"Position: Video as the New Language for Real-World Decision Making","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"","doi":"","generation_model":"多种示例：Transformer-XL + VQ-VAE / MaskGIT；pixel-space 3D U-Net diffusion；Genie 风格 masked dynamics","auxiliary_models":"依示例分别采用动作 tokenizer / latent action model；无单一统一 planner","conditions":"文本、观察帧、离散动作或 latent actions，随示例而变","training_datasets":"Position paper 汇集多种例子；正文明确有 Open X-Embodiment 训练的生成示例；其他实例的完整语料与划分未核实","evaluation_datasets":"规划、机器人、游戏与 simulation 的说明性示例","evidence_location":"§3–5；Appendix A.1–A.3，PDF p.15","training_status":"立场论文，含多个训练模型示例","annotation_notes":"不将文中所有引用模型或数据集合并为本文一个训练配置。","arxiv_id":"未核实","evidence_url":"https://raw.githubusercontent.com/mlresearch/v235/main/assets/yang24z/yang24z.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICML-2025-0008","venue":"ICML","year":2025,"title":"AdaWorld: Learning Adaptable World Models with Latent Actions","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=QQegZj99sk","doi":"","generation_model":"Stable Video Diffusion 初始化 autoregressive world model；另验证 iVideoGPT 变体","auxiliary_models":"自监督 latent action autoencoder；下游 action adaptation / MPC","conditions":"历史帧 + latent actions；可从 demonstration video 提取动作，或映射目标域真实动作","training_datasets":"Open X-Embodiment、Ego4D、Something-Something-v2、MiraData；Gym Retro / Procgen 的 1,016 environments 自动采样，合计约 2B frames；下游按环境适配","evaluation_datasets":"LIBERO、SSv2 action transfer；RoboDesk / Robosuite planning；BAIR、Minecraft、nuScenes 等适配实验","evidence_location":"§3–4，PDF pp.5–9；Appendix A–B，pp.15–17","training_status":"latent action 预训练 + world model 预训练 + 下游适配","annotation_notes":"iVideoGPT 是验证方法通用性的额外 backbone，主要实现基于 SVD。","arxiv_id":"2503.18938","evidence_url":"https://arxiv.org/pdf/2503.18938","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICML-2025-0897","venue":"ICML","year":2025,"title":"Empowering World Models with Reflection for Embodied Video Prediction","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=onumui0nHi","doi":"","generation_model":"DynamiCrafter 视频生成器，EVA-Generator / EVA","auxiliary_models":"ChatUniVi VLM；generation adapter / ensemble-LoRA；GPT-4o / GPT-4V 构造标注与评测","conditions":"历史视频 / 图像 + embodied task question；RoG reasoning 输出进一步条件化视频预测","training_datasets":"EVA-Instruct 500K：来源含 OXE / RT-1、RoboVQA、Ego4D、Ego-Exo4D、CALVIN；VLM instruction stage 另含 MIMIC-IT、LLaVA instruction、VideoChatGPT","evaluation_datasets":"EVA-Bench，in-domain / OOD；How-to、Action-Description、Finish-Think、Next-Step","evidence_location":"§5–6，PDF pp.5–7；Appendix A.6–A.7，pp.15–16","training_status":"多阶段 VLM / generator / adapter 训练","annotation_notes":"正文简述 adapter-only，附录另写 ensemble-LoRA 版本；按不同变体记录，避免混成单一配置。","arxiv_id":"未核实","evidence_url":"https://raw.githubusercontent.com/mlresearch/v267/main/assets/chi25b/chi25b.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2024-0388","venue":"NeurIPS","year":2024,"title":"Closed-Loop Visuomotor Control with Generative Expectation for Robotic Manipulation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/fad8962279154544ed69bb63eb14d677-Abstract-Conference.html","doi":"","generation_model":"CLOVER 自训 RGB-D video diffusion；Imagen-style U-Net + causal temporal attention","auxiliary_models":"RAFT ContextNet 作训练 flow regularizer；feedback policy 比较 VC-1 / CLIP / DINO visual encoders 与 ViT-Small depth encoder","conditions":"当前 RGB-D 观察 + 文本任务；生成 future subgoals，误差反馈触发重规划","training_datasets":"CALVIN（policy 明确 ABC training split）；真机 50 collected demonstrations 微调 diffusion planner","evaluation_datasets":"CALVIN、真机长程 manipulation / 扰动恢复","evidence_location":"Appendix B.1–B.2，PDF pp.17–18","training_status":"video planner + feedback policy 训练","annotation_notes":"真实环境无 depth 数据，真机模型仅生成 RGB；flow estimation 模块测试时丢弃。","arxiv_id":"未核实","evidence_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/fad8962279154544ed69bb63eb14d677-Paper-Conference.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2024-1367","venue":"NeurIPS","year":2024,"title":"Compositional 3D-aware Video Generation with LLM Director","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/edbeca7811f9365c924c72a8a9bce83a-Abstract-Conference.html","doi":"","generation_model":"LucidDreamer 生成 3D scene、HumanGaussian 生成 humanoid、Motion-X 动作；Stable Diffusion 提供 SDS prior","auxiliary_models":"LLM / multimodal LLM director 负责分解与粗尺度 / 位置 / 轨迹规划；精确 checkpoint 未核实","conditions":"文本 → 分解实体、3D assets、scale / location / motion trajectory 与 camera controls","training_datasets":"未训练新的通用视频 backbone；使用预训练专家并对每个场景执行 SDS / 位置尺度优化；Motion-X 在此为动作资源","evaluation_datasets":"compositional 3D-aware video prompts 与视觉 / 可控性评测","evidence_location":"§5，PDF p.6；Appendix A，p.15","training_status":"预训练专家 + 每场景优化","annotation_notes":"每场景优化不同于训练一个通用生成模型；不把专家原训练语料混入本文新增训练数据。","arxiv_id":"未核实","evidence_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/edbeca7811f9365c924c72a8a9bce83a-Paper-Conference.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"auxiliary_models","classification_review":""},{"id":"NeurIPS-2024-2391","venue":"NeurIPS","year":2024,"title":"iVideoGPT: Interactive VideoGPTs are Scalable World Models","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/7dbb5bfab324e3b86af9bd0df15498dd-Abstract-Conference.html","doi":"","generation_model":"iVideoGPT autoregressive Transformer + compressive VQGAN tokenizer；包含 436M scaling variant","auxiliary_models":"action / reward token prediction；用于 MPC 与 model-based RL","conditions":"视觉 context frames、动作、奖励；goal-conditioned 变体另提供末帧目标","training_datasets":"tokenizer：OXE 35 subsets + SSv2；Transformer：仅 OXE；下游 BAIR、RoboNet、VP² 数据适配；RL 另使用环境交互","evaluation_datasets":"action-conditioned video prediction、VP² visual planning、model-based RL","evidence_location":"Appendix A.1–A.2，PDF pp.18–20；下游实验章节","training_status":"预训练 + 下游 fine-tuning / RL","annotation_notes":"SSv2 用于 tokenizer 而非本文默认 Transformer 预训练，已明确拆开。","arxiv_id":"未核实","evidence_url":"https://proceedings.neurips.cc/paper_files/paper/2024/file/7dbb5bfab324e3b86af9bd0df15498dd-Paper-Conference.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2025-2377","venue":"NeurIPS","year":2025,"title":"TalkCuts: A Large-Scale Dataset for Multi-Shot Human Speech Video Generation","reasoning_role":"R1","topic":"评测 / 数据 / 分析 / 安全","url":"https://openreview.net/forum?id=4a0w7AkrY7","doi":"","generation_model":"Orator multimodal video generation baseline；具体 checkpoint 未核实","auxiliary_models":"LLM director 规划 camera transitions / gestures / vocal modulation；LLM 型号未核实","conditions":"pose-guided 或 audio-driven；内部有分镜、动作与声音计划","training_datasets":"TalkCuts：164K clips、超过 500 小时、10K+ identities；摘要明确训练使用 TalkCuts，具体 train split 未核实","evaluation_datasets":"TalkCuts multi-shot speech video generation","evidence_location":"官方录用摘要；未获得可读正文","training_status":"有训练，细节未核实","annotation_notes":"164K / 500 小时为数据集总规模，不是已确认的 training split 规模。","arxiv_id":"未核实","evidence_url":"https://openreview.net/forum?id=4a0w7AkrY7","evidence_level":"仅官方摘要","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model；auxiliary_models；training_datasets","classification_review":""},{"id":"NeurIPS-2025-2664","venue":"NeurIPS","year":2025,"title":"UniTransfer: Video Concept Transfer via Progressive Spatio-Temporal Decomposition","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openreview.net/forum?id=tsEHIfv2lM","doi":"","generation_model":"CogVideoX-based dual-to-single-stream DiT；精确参数规模未核实","auxiliary_models":"foreground / background / motion decomposition；无已确认外部 LLM planner","conditions":"目标视频 + 参考主体 / 服装 / 背景图像，或 reference motion；文本 embeddings","training_datasets":"OpenVid 自监督预训练；TikTok + UBC Fashion 人体任务微调；OpenAnimal 动物任务微调","evaluation_datasets":"human / animal / object / background / motion transfer","evidence_location":"§3–4；Appendix B，PDF p.14","training_status":"self-supervised pretraining + supervised fine-tuning","annotation_notes":"所读版本标题为 Spatial and Timestep Decomposition。正文称 OpenVid 约 12M，与通常 OpenVid-1M 命名有出入，故不将该规模作为确认值；原严格 R1 标签需复审，主文主要是分解编辑而非明确 reasoning。","arxiv_id":"2509.21086","evidence_url":"https://arxiv.org/pdf/2509.21086","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":"待复审：严格 R1 依据偏弱"},{"id":"NeurIPS-2025-4954","venue":"NeurIPS","year":2025,"title":"MindJourney: Test-Time Scaling with World Models for Spatial Reasoning","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=L2W4wQsNkY","doi":"","generation_model":"Search World Model：Wan2.2-TI2V-5B + ReCamMaster 风格 camera conditioning","auxiliary_models":"MindJourney 接 GPT-4o、GPT-4.1、InternVL3-14B、OpenAI o1 等 VLM 进行探索 / 空间问答","conditions":"场景图像 + 空间问题；VLM 输出 primitive camera actions / camera extrinsics 作为 world-model 条件","training_datasets":"SWM camera encoder / projector：Habitat 在 HM3D 生成的 50K navigation clips + RealEstate10K + DL3DV-10K；VLM 无新增微调","evaluation_datasets":"SAT-Synthesized、SAT-Real","evidence_location":"§3.4–4.1，PDF pp.6–7；Appendix A.2，p.15","training_status":"world-model 控制模块训练；VLM test-time scaling","annotation_notes":"官方摘要称 SpatialNavigator；所读更新版称 MindJourney。without fine-tuning 指 VLM 推理环节，不表示 SWM 不训练。","arxiv_id":"2507.12508","evidence_url":"https://arxiv.org/pdf/2507.12508","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"NeurIPS-2025-5490","venue":"NeurIPS","year":2025,"title":"Learning 3D Persistent Embodied World Models","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openreview.net/forum?id=XTTbzC7O2T","doi":"","generation_model":"CogVideoX，加入 RGB-D / action channels 与 3D memory blocks；具体 checkpoint 规模未核实","auxiliary_models":"persistent 3D feature map；下游 planning / policy learning","conditions":"当前观察、动作 / camera embeddings、历史累积 3D map；输出 RGB-D futures","training_datasets":"Habitat + HM3D 自采约 50K trajectories，每条最多 500 steps；场景分 train / test；先适配 backbone，再冻结 backbone 训练 memory blocks","evaluation_datasets":"held-out HM3D scenes；long-horizon persistence、planning 与 policy learning","evidence_location":"§3.3–4.1，PDF pp.5–6；Supplement §1.2，p.11","training_status":"两阶段 world-model / memory 训练","annotation_notes":"约 1,000 scenes 是采集场景范围，不等于全部用于 training。","arxiv_id":"2505.05495","evidence_url":"https://arxiv.org/pdf/2505.05495","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"CVPR-2024-0659","venue":"CVPR","year":2024,"title":"Dysen-VDM: Empowering Dynamics-aware Text-to-Video Diffusion with LLMs","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2024/html/Fei_Dysen-VDM_Empowering_Dynamics-aware_Text-to-Video_Diffusion_with_LLMs_CVPR_2024_paper.html","doi":"","generation_model":"Latent VDM autoencoder + 3D U-Net；加入 RGTrm dynamic scene graph encoder","auxiliary_models":"ChatGPT（GPT-3.5 / GPT-4 比较）执行 Dysen action scheduling / scene graph enrichment","conditions":"文本 → 有时间顺序的 actions → enriched dynamic scene graph + 文本","training_datasets":"WebVid 预训练 latent VDM 与 DSG-conditioned VDM；另报告 UCF-101 / MSR-VTT 的直接 fine-tuning 设置","evaluation_datasets":"UCF-101、MSR-VTT；zero-shot 与 fine-tuned 两种协议","evidence_location":"§4–5，PDF pp.5–6；GPT comparison，p.8","training_status":"video / graph encoder 训练；LLM in-context reasoning","annotation_notes":"不能将 zero-shot evaluation 解读为整套方法从未训练。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2024/papers/Fei_Dysen-VDM_Empowering_Dynamics-aware_Text-to-Video_Diffusion_with_LLMs_CVPR_2024_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2024-1231","venue":"CVPR","year":2024,"title":"Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/CVPR2024/html/Wang_Driving_into_the_Future_Multiview_Visual_Forecasting_and_Planning_with_CVPR_2024_paper.html","doi":"","generation_model":"Stable Diffusion 初始化 multiview temporal diffusion（Drive-WM）；具体 SD 版本未核实","auxiliary_models":"image-based reward；兼容 VAD 等 driving planners","conditions":"多视角历史图像、文本描述；按变体给 ego actions 或 layouts / 邻接参考视角","training_datasets":"nuScenes：700 training scenes；生成 counterfactual driving videos 另用于 planner fine-tuning","evaluation_datasets":"nuScenes 150 validation scenes；multiview generation 与 trajectory planning","evidence_location":"§3–5.1，PDF pp.4–7","training_status":"world-model 训练；另有 planner 数据增强实验","annotation_notes":"VAD 是下游 planner，不是视频 backbone。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2024/papers/Wang_Driving_into_the_Future_Multiview_Visual_Forecasting_and_Planning_with_CVPR_2024_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"CVPR-2025-0149","venue":"CVPR","year":2025,"title":"BlobGEN-Vid: Compositional Text-to-Video Generation with Blob Video Representations","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openaccess.thecvf.com/content/CVPR2025/html/Feng_BlobGEN-Vid_Compositional_Text-to-Video_Generation_with_Blob_Video_Representations_CVPR_2025_paper.html","doi":"","generation_model":"VideoCrafter2（U-Net）与 CogVideoX-5B（DiT）两个 backbone","auxiliary_models":"GPT-4o 在完整 T2V pipeline 中做 blob layout planning；LLaVA-NeXT 产生训练 object captions","conditions":"全局文本 + blob geometry / per-object captions / object trajectories；可做多视角场景生成","training_datasets":"约 1M dense-blob annotated videos，来自 OpenVid-1M、VidGEN-1M、少量 HD-VILA；indoor variant 使用 BlobGEN-3D 处理的 ScanNet++；CogVideoX 变体只用约一半主数据","evaluation_datasets":"layout controllability、compositional T2V、多视角 ScanNet++","evidence_location":"§4–5，PDF pp.5–8","training_status":"冻结 backbone，训练新增 control layers","annotation_notes":"数据规模随 backbone / ablation 改变，不能统一称所有模型均训练 1M。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Feng_BlobGEN-Vid_Compositional_Text-to-Video_Generation_with_Blob_Video_Representations_CVPR_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2025-1103","venue":"CVPR","year":2025,"title":"PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/CVPR2025/html/Xue_PhyT2V_LLM-Guided_Iterative_Self-Refinement_for_Physics-Grounded_Text-to-Video_Generation_CVPR_2025_paper.html","doi":"","generation_model":"CogVideoX（2B / 5B）、OpenSora 1.2、VideoCrafter","auxiliary_models":"GPT-4o reasoning / prompt refinement；Tarsier video captioner；VideoCon-Physics evaluator","conditions":"文本；生成视频被转成反馈供迭代 CoT / step-back prompt refinement","training_datasets":"无新增模型训练；不需要额外训练数据","evaluation_datasets":"VideoPhy、PhyGenBench；另有 VBench","evidence_location":"§3–4，PDF pp.4–6","training_status":"Training-free iterative refinement","annotation_notes":"物理 benchmark prompts 是推理 / 测试输入，不是训练语料。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Xue_PhyT2V_LLM-Guided_Iterative_Self-Refinement_for_Physics-Grounded_Text-to-Video_Generation_CVPR_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"CVPR-2025-2277","venue":"CVPR","year":2025,"title":"Navigation World Models","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/CVPR2025/html/Bar_Navigation_World_Models_CVPR_2025_paper.html","doi":"","generation_model":"NWM Conditional DiT-XL，约 1B；Stable Diffusion VAE tokenizer","auxiliary_models":"simulation-based planning / external policy trajectory ranking","conditions":"默认 4 个历史观察帧 + navigation actions / time shift；可给 goal observation 用于规划","training_datasets":"SCAND、TartanDrive、RECON、HuRoN；另使用无动作标签 Ego4D 视频","evaluation_datasets":"已知环境 held-out trajectories；GO Stanford 作为未知环境","evidence_location":"§3–4.1，PDF pp.4–5","training_status":"action-conditioned world model 训练","annotation_notes":"GO Stanford 在本文为 unknown-environment evaluation，不能因 baseline 训练用过而列入 NWM training。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/CVPR2025/papers/Bar_Navigation_World_Models_CVPR_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICCV-2025-0423","venue":"ICCV","year":2025,"title":"Aether: Geometric-Aware Unified World Modeling","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Zhu_Aether_Geometric-Aware_Unified_World_Modeling_ICCV_2025_paper.html","doi":"","generation_model":"CogVideoX-5B-I2V 初始化 Aether；新增 depth / raymap input-output channels","auxiliary_models":"自动 synthetic camera / depth annotation pipeline；无独立文本 planner","conditions":"观察帧、目标帧、camera trajectories / raymaps；可输入视频用于 reconstruction；文本固定为空","training_datasets":"按照 DA-V / The-Matrix 路线采集并标注的 synthetic RGB-D videos；精确数据源清单 / 规模未核实","evaluation_datasets":"synthetic in-domain / out-domain prediction；Sintel、BONN、KITTI reconstruction 等 real zero-shot tests","evidence_location":"§2–3.4，PDF pp.2–5；§4–5，pp.6–7","training_status":"synthetic post-training","annotation_notes":"只用 synthetic 指本文 post-training；CogVideoX 原始预训练并非由此证明未用真实视频。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhu_Aether_Geometric-Aware_Unified_World_Modeling_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICCV-2025-0426","venue":"ICCV","year":2025,"title":"VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Yang_VLIPP_Towards_Physically_Plausible_Video_Generation_with_Vision_and_Language_ICCV_2025_paper.html","doi":"","generation_model":"Go-with-the-Flow image-to-video model","auxiliary_models":"GPT-4o coarse physical motion planner；Grounded-SAM2；FLUX 生成部分评测首帧","conditions":"图像 + 文本 → 物理对象 boxes / rough motion trajectories；带噪的 motion control 驱动 I2V","training_datasets":"所读方法与实验采用现成模型执行推理，未发现本文新增权重训练；底座原始训练数据未核实","evaluation_datasets":"PhyGenBench（FLUX 首帧适配）；Physics-IQ","evidence_location":"§3–4.2，PDF pp.4–6","training_status":"现成模型推理；新增训练未见明确报告","annotation_notes":"不将 Go-with-the-Flow 自身训练语料作为 VLIPP 的新增训练数据。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Yang_VLIPP_Towards_Physically_Plausible_Video_Generation_with_Vision_and_Language_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICCV-2025-1440","venue":"ICCV","year":2025,"title":"Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis","reasoning_role":"R1","topic":"质量 / 物理 / 推理 / 训练优化","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Li_Causal-Entity_Reflected_Egocentric_Traffic_Accident_Video_Synthesis_ICCV_2025_paper.html","doi":"","generation_model":"Causal-VidSyn 3D U-Net video diffusion + CTS / CTG modules；具体预训练 checkpoint 未核实","auxiliary_models":"accident reason answering head；gaze-guided causal entity selection","conditions":"训练：事故 cause / prevention descriptions、视频、driver fixation；测试：视频或文本 prompts，不需 gaze","training_datasets":"MM-AU 的 9,727 事故视频中采 6,492 near-crash-to-crash clips；Drive-Gaze 提供 driver fixation 监督；三个 progressive training stages","evaluation_datasets":"DADA-2000 editing / T2V；BDD-A normal-to-accident generation","evidence_location":"§4–5.2，PDF pp.5–6","training_status":"三阶段 video / causal grounding 训练","annotation_notes":"CogVideoX-2B、Latte 是比较模型，不是本文默认 backbone。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Li_Causal-Entity_Reflected_Egocentric_Traffic_Accident_Video_Synthesis_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"ICCV-2025-1517","venue":"ICCV","year":2025,"title":"DrivingGPT: Unifying Driving World Modeling and Planning with Multi-modal Autoregressive Transformers","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Chen_DrivingGPT_Unifying_Driving_World_Modeling_and_Planning_with_Multi-modal_Autoregressive_ICCV_2025_paper.html","doi":"","generation_model":"DrivingGPT：从头训练 110M autoregressive Transformer；LlamaGen visual tokenizer","auxiliary_models":"同一个模型联合预测 image / action tokens；无独立 LLM planner","conditions":"历史 front-view images 与 driving actions 的交错 token sequence","training_datasets":"NAVSIM navtrain；另对比均匀采样 nuPlan 651K sequences 与 NAVSIM train 104K sequences 的数据配方","evaluation_datasets":"nuPlan / NAVSIM navtest；video prediction 与 driving planning","evidence_location":"§4.1–4.5，PDF pp.5–8","training_status":"Transformer from scratch；使用预训练 tokenizer","annotation_notes":"SVD 是对照基线，不能标作 DrivingGPT 的 backbone。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Chen_DrivingGPT_Unifying_Driving_World_Modeling_and_Planning_with_Multi-modal_Autoregressive_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICCV-2025-1556","venue":"ICCV","year":2025,"title":"Long-Context State-Space Video World Models","reasoning_role":"R3","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Po_Long-Context_State-Space_Video_World_Models_ICCV_2025_paper.html","doi":"","generation_model":"block-wise Mamba state-space video diffusion + dense local attention；diffusion forcing","auxiliary_models":"SSM temporal memory；没有单独的外部 LLM","conditions":"历史视觉观察 + agent actions / positions；不提供 maze ground-truth layout","training_datasets":"Memory Maze trajectories；TECO Minecraft 200K gameplay trajectories（每条 1,501 action-observation pairs）的 training split","evaluation_datasets":"Memory Maze / Minecraft spatial retrieval、long-context reasoning","evidence_location":"§3–5.1，PDF pp.3–6","training_status":"action-conditioned world-model 训练","annotation_notes":"200K 是 TECO 数据总体规模；具体训练 / 测试划分数量未核实。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Po_Long-Context_State-Space_Video_World_Models_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICCV-2025-1565","venue":"ICCV","year":2025,"title":"VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation","reasoning_role":"R1","topic":"Editing / 修改已有视频","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Yu_VEGGIE_Instructional_Editing_and_Reasoning_Video_Concepts_with_Grounded_Generation_ICCV_2025_paper.html","doi":"","generation_model":"MagicBrush 初始化 SD1.5 UNet，加 AnimateDiff 风格 temporal layers","auxiliary_models":"LLaVA-OneVision-7B（Qwen2 / SigLIP）+ grounded task queries / alignment network","conditions":"原始视频 + 自然语言编辑 / grounding / reasoning 指令；MLLM 生成 frame-specific task tokens","training_datasets":"Stage 1：Seed-Data-Edit、LISA、gRefCoCo、PhraseCut、EraseDraw、MagicBrush、SmartEdit；Stage 2：ROVI、VPLM、GroundMoRe、RVoS、MeViS、InstructV2V、自建 VEG-Edit；Table 1 列约 3.438M image instructions、136.1K video instructions","evaluation_datasets":"instructional video editing、referring grounding、reasoning segmentation；各任务 held-out split","evidence_location":"主文 §3.2–3.3 / Table 1，PDF pp.4–5；Supplement §6.1，p.1","training_status":"两阶段 image-to-video curriculum，端到端训练","annotation_notes":"正文叙述 video 总量 133.9K 与 Table 1 的 136.1K 不一致，此处明确采用表格数；附录与主文分别链接。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/supplemental/Yu_VEGGIE_Instructional_Editing_ICCV_2025_supplemental.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Yu_VEGGIE_Instructional_Editing_and_Reasoning_Video_Concepts_with_Grounded_Generation_ICCV_2025_paper.pdf","unresolved_fields":"","classification_review":""},{"id":"ICCV-2025-1614","venue":"ICCV","year":2025,"title":"MotionAgent: Fine-grained Controllable Video Generation via Motion Field Agent","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Liao_MotionAgent_Fine-grained_Controllable_Video_Generation_via_Motion_Field_Agent_ICCV_2025_paper.html","doi":"","generation_model":"Stable Video Diffusion 冻结，训练 optical flow adapter","auxiliary_models":"GPT-4o motion-field agent；3D analytical optical flow composition；可选 rethinking","conditions":"初始图像 + 文本 → object trajectories / camera extrinsics → dense optical flow","training_datasets":"adapter 使用视频及提取的 optical flow 监督；所读主文未确认训练数据集名称，保留未核实","evaluation_datasets":"VBench、motion-focused VBench subset；30 images 的 user-study 输入集","evidence_location":"§3–4.1，PDF pp.3–5；§4.2.3，p.7","training_status":"flow adapter 训练；LLM 推理","annotation_notes":"不用高质量 video-image-text pairs 不等于不用训练数据；30 是 user-study 图像数，不是参与人数。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Liao_MotionAgent_Fine-grained_Controllable_Video_Generation_via_Motion_Field_Agent_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"ICCV-2025-1714","venue":"ICCV","year":2025,"title":"Epona: Autoregressive Diffusion World Model for Autonomous Driving","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Zhang_Epona_Autoregressive_Diffusion_World_Model_for_Autonomous_Driving_ICCV_2025_paper.html","doi":"","generation_model":"Epona 2.5B：1.3B spatiotemporal Transformer + 1.2B next-frame diffusion + 50M trajectory diffusion；temporal-aware DCAE","auxiliary_models":"video / trajectory 联合 world model；无独立 LLM planner","conditions":"历史驾驶图像与轨迹；默认 10 past frames，autoregressive future prediction","training_datasets":"nuPlan 公开视频 + nuScenes 700 scenes；world model 从头训练，DCAE 基于预训练组件适配","evaluation_datasets":"nuPlan test、nuScenes validation；nuScenes / NAVSIM trajectory planning","evidence_location":"§3–4.1，PDF pp.5–6","training_status":"world model from scratch + autoencoder 适配","annotation_notes":"NAVSIM 在此段为 planning benchmark，训练数据主配置明确写 nuPlan / nuScenes。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhang_Epona_Autoregressive_Diffusion_World_Model_for_Autonomous_Driving_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"ICCV-2025-1897","venue":"ICCV","year":2025,"title":"IRASim: A Fine-Grained World Model for Robot Manipulation","reasoning_role":"R2","topic":"动作预测 / 交互 World Models","url":"https://openaccess.thecvf.com/content/ICCV2025/html/Zhu_IRASim_A_Fine-Grained_World_Model_for_Robot_Manipulation_ICCV_2025_paper.html","doi":"","generation_model":"IRASim diffusion Transformer + frame-level action conditioning；小数据 policy 实验初始化 OpenSora","auxiliary_models":"diffusion policy；ResNet50 IoU value model 用于 PushT planning","conditions":"历史图像 + 逐帧机器人 action trajectories","training_datasets":"RT-1、Bridge、Language-Table、RoboNet 的训练划分；LIBERO / PushT 实验用 expert demos + policy post-trained rollouts（含成功 / 失败）","evaluation_datasets":"四套真实机器人视频测试；LIBERO policy evaluation；PushT model-based planning","evidence_location":"§4.1–4.3，PDF pp.4–7","training_status":"world model / value training；policy rollout 数据增强","annotation_notes":"OpenSora 初始化明确用于小数据 policy evaluation / planning，未据此断言所有生成实验均用同一 checkpoint。","arxiv_id":"未核实","evidence_url":"https://openaccess.thecvf.com/content/ICCV2025/papers/Zhu_IRASim_A_Fine-Grained_World_Model_for_Robot_Manipulation_ICCV_2025_paper.pdf","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"AAAI-2025-1293","venue":"AAAI","year":2025,"title":"Modular-Cam: Modular Dynamic Camera-view Video Generation with LLM","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://doi.org/10.1609/aaai.v39i6.32681","doi":"10.1609/aaai.v39i6.32681","generation_model":"SD1.5 + temporal Transformer；CamOperator LoRA 与 AdaControlNet","auxiliary_models":"chatgpt-3.5-turbo-1106 LLM-Director；底座加载 DreamBooth LoRA","conditions":"文本 → scenes / camera transitions；六种相机运动模块；可选首帧图像","training_datasets":"WebVid-10M 训练 temporal layers；约 50 条按运动模式增强的 videos 训练 CamOperator；WebVid-10M 100K subset 训练 AdaControlNet","evaluation_datasets":"多场景 / 动态相机文本生成实验；具体测试集名称未核实","evidence_location":"Experiment Setup，PDF p.5；Appendix implementation / algorithm，pp.9–10","training_status":"分模块训练","annotation_notes":"约 50 条为文中 CamOperator 合成数据描述；未推断为每类各 50。","arxiv_id":"2504.12048","evidence_url":"https://arxiv.org/pdf/2504.12048","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"AAAI-2025-2219","venue":"AAAI","year":2025,"title":"DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://doi.org/10.1609/aaai.v39i10.33130","doi":"10.1609/aaai.v39i10.33130","generation_model":"Stable Video Diffusion + UniMVM；SD2.1 ControlNet 生成 HDMap","auxiliary_models":"GPT-3.5 在 text-to-script 数据上微调，用轨迹函数库生成 agent trajectories","conditions":"用户文本 → agent trajectories → HDMap / projected 3D boxes；可附多视角首帧或单视角视频","training_datasets":"nuScenes 700 training scenes，约 1M frames；构造 text-to-script 数据微调 GPT-3.5；HDMap 与 video generator 分别训练","evaluation_datasets":"nuScenes 150 validation scenes；生成视频用于 detection / tracking 数据增强","evidence_location":"§4.1，PDF pp.9–10；Table 5，p.14","training_status":"LLM / HDMap / video 多模块训练","annotation_notes":"参考 arXiv 正文版本；SVD 全参数 fine-tuning，HDMap 仅 ControlNet 参数可训练。","arxiv_id":"2403.06845","evidence_url":"https://arxiv.org/pdf/2403.06845","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"AAAI-2026-3029","venue":"AAAI","year":2026,"title":"DreamRunner: Fine-Grained Compositional Story-to-Video Generation with Retrieval-Augmented Motion Adaptation","reasoning_role":"R1","topic":"长视频 / 多镜头 / 记忆","url":"https://doi.org/10.1609/aaai.v40i13.38022","doi":"10.1609/aaai.v40i13.38022","generation_model":"CogVideoX-2B（story）；另用 CogVideoX-5B 做 compositional T2V","auxiliary_models":"GPT-4o coarse / fine story planning；spatiotemporal region attention","conditions":"故事文本、character reference images；逐实体 motions / boxes 与检索得到的 motion examples","training_datasets":"test-time adaptation：top-ranked retrieved motion videos 训练 temporal / spatial LoRA；参考图像训练 subject prior；检索视频库的具体 dataset 名称未核实","evaluation_datasets":"DreamStorySet、T2VComBench","evidence_location":"§3–4.1，PDF pp.3–6；Appendix A.1–A.2，pp.10–11","training_status":"test-time LoRA fine-tuning","annotation_notes":"检索素材确实用于更新 LoRA，因此这里不能标成 training-free RAG。","arxiv_id":"2411.16657","evidence_url":"https://arxiv.org/pdf/2411.16657","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"AAAI-2026-3148","venue":"AAAI","year":2026,"title":"ManipDreamer3D: Synthesizing Plausible Robotic Manipulation Video with Occupancy-aware 3D Trajectory","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://doi.org/10.1609/aaai.v40i8.37595","doi":"10.1609/aaai.v40i8.37595","generation_model":"两个实现：Stable Video Diffusion；CogVideoX-5B","auxiliary_models":"VGGT 3D reconstruction；Qwen-VL grounding；fine-tuned YOLO gripper detector；occupancy-aware trajectory optimizer","conditions":"初始图像 + 文本指令 → 3D occupancy 与优化后的 end-effector trajectory；latent editing 控制视频","training_datasets":"Bridge V1 + Bridge V2，经 curation 得 8.7K valid episodes，按 9:1 train / test 划分；辅助 detector 的独立训练数据未核实","evaluation_datasets":"held-out curated Bridge manipulation episodes","evidence_location":"3D Trajectory Data Curation，PDF p.4；Experiment Setups，p.6","training_status":"trajectory-conditioned generator 训练","annotation_notes":"8.7K 为处理后总量，不是 training split 数量。","arxiv_id":"2509.05314","evidence_url":"https://arxiv.org/pdf/2509.05314","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"AAAI-2026-3449","venue":"AAAI","year":2026,"title":"FantasyHSI: Video-Generation-Centric 4D Human Synthesis in Any Scene Through a Graph-Based Multi-Agent Framework","reasoning_role":"R1","topic":"人物动画 / Talking Head / HOI","url":"https://doi.org/10.1609/aaai.v40i10.37758","doi":"10.1609/aaai.v40i10.37758","generation_model":"Wan2.1-I2V-14B","auxiliary_models":"Gemini-2.5-Pro graph-based multi-agents：navigator / planner / critic","conditions":"场景 / 角色观察、长程文本目标；graph plan 与逐步动作指导；视频后处理为 4D HSI","training_datasets":"自建 10,000 preference pairs 做 DPO；候选来自 Veo、Hunyuan-Video、Runway、Kling，由专业标注者按指令 / 动作 / 形体 / 场景一致性标偏好","evaluation_datasets":"SceneBench：TRUMANS 与 Sketchfab 场景组成的 20-scene benchmark","evidence_location":"§3.3–4.2，PDF p.5","training_status":"generator DPO；agent 推理","annotation_notes":"without paired data 不表示无训练；SceneBench 是评测场景集合，不是 DPO preference dataset。","arxiv_id":"2509.01232","evidence_url":"https://arxiv.org/pdf/2509.01232","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"AAAI-2026-3563","venue":"AAAI","year":2026,"title":"GENMAC: Compositional Text-to-Video Generation with Multi-Agent Collaboration","reasoning_role":"R1","topic":"条件控制 / 定制 / 几何一致","url":"https://doi.org/10.1609/aaai.v40i7.37418","doi":"10.1609/aaai.v40i7.37418","generation_model":"VideoCrafter2","auxiliary_models":"GPT-4o 多 agent：Design / Verification / Suggestion / Redesign","conditions":"复杂文本 → frame-wise layouts、修订 prompts 与 bbox guidance scale；根据已生成视频迭代修正","training_datasets":"无新增模型训练；现成生成器上用 LVD-style attention guidance","evaluation_datasets":"compositional text-to-video benchmarks；具体 benchmark 名称未在本次配置摘录中核实","evidence_location":"§3–4.1，PDF pp.4–5；Appendix B.1，p.12","training_status":"Training-free multi-agent inference","annotation_notes":"与文中比较的 17 个视频模型区分，主生成 backbone 是 VideoCrafter2。","arxiv_id":"2412.04440","evidence_url":"https://arxiv.org/pdf/2412.04440","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"MM-2025-0208","venue":"MM","year":2025,"title":"SpA2V: Harnessing Spatial Auditory Cues for Audio-driven Spatially-aware Video Generation","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://doi.org/10.1145/3746027.3755705","doi":"10.1145/3746027.3755705","generation_model":"Stable Diffusion + pretrained AnimateDiff temporal modules + MIGC grounding modules","auxiliary_models":"Gemini-2.0-Flash video planner；CLAP kNN 检索 3 个 in-context examples；LLaVA-OneVision 辅助标注","conditions":"输入 stereo audio；MLLM 推断对象、位置与运动形成 VSL / global-local captions，控制视频生成","training_datasets":"无新增权重训练，组合已有模块；in-context candidate examples 用于推理检索","evaluation_datasets":"AVLBench：7,274 test samples，stationary instruments / translating vehicles","evidence_location":"§3–4.1，PDF pp.3–6；Appendix A","training_status":"Training-free + in-context retrieval","annotation_notes":"AVLBench 的评测样本及示例库不能填成 generator training dataset。","arxiv_id":"2508.00782","evidence_url":"https://arxiv.org/pdf/2508.00782","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"SIGGRAPH-2026-0075","venue":"SIGGRAPH","year":2026,"title":"MAViD: A Multimodal Framework for Audio-Visual Dialogue Understanding and Generation","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://doi.org/10.1145/3799902.3811223","doi":"10.1145/3799902.3811223","generation_model":"Creator：autoregressive audio model + Wan DiT / VAE；higgs-audio tokenizer；Creator 合计约 8B","auxiliary_models":"Conductor：Qwen2.5-Omni Thinker；Gemini-2.5-Pro 生成数据标注","conditions":"任意组合文本 / 图像 / 音频 / 视频 query → speech 与 motion instructions；可选 identity reference image；历史 audio-video clips","training_datasets":"Conductor：混合 multimodal QA / human-centered instruction 数据，精确清单未核实；Creator audio stage：自采超过 10K 小时网络音频；joint stage：近 1M audio-video pairs（影片、talking body、演讲、互动）","evaluation_datasets":"多模态理解 benchmarks 与独立 audio-video generation evaluation","evidence_location":"§3.2–4.1，PDF pp.4–6","training_status":"三阶段 Conductor / audio AR / joint AV 训练","annotation_notes":"CosyVoice2 + Wan-S2V 是比较基线，不是本文 Creator 主配置；Wan 精确版本未核实。","arxiv_id":"2512.03034","evidence_url":"https://arxiv.org/pdf/2512.03034","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"training_datasets","classification_review":""},{"id":"SIGGRAPH-Asia-2024-0116","venue":"SIGGRAPH-Asia","year":2024,"title":"Anim-Director: A Large Multimodal Model Powered Agent for Controllable Animation Video Generation","reasoning_role":"R1","topic":"人物动画 / Talking Head / HOI","url":"https://doi.org/10.1145/3680528.3687688","doi":"10.1145/3680528.3687688","generation_model":"Midjourney 图像工具；Pika 视频工具；API 具体 snapshot 未核实","auxiliary_models":"GPT-4 LMM director / self-reflection；SAM 区域分割辅助图像修正","conditions":"简短故事 / 指令 → director script、角色与背景参考、scene images + video prompts","training_datasets":"无新增权重训练；通过现成 API 生成、评估、选择与修改","evaluation_datasets":"TinyStories 中选 100 narratives 作为评测输入","evidence_location":"§1 / framework，PDF pp.2–3；§3–4.1，pp.5–6","training_status":"Training-free agent orchestration","annotation_notes":"TinyStories 在本文是测试故事来源，不是训练生成器的数据。","arxiv_id":"2408.09787","evidence_url":"https://arxiv.org/pdf/2408.09787","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""},{"id":"SIGGRAPH-Asia-2025-0111","venue":"SIGGRAPH-Asia","year":2025,"title":"AniMaker: Multi-Agent Animated Storytelling with MCTS-Driven Clip Generation","reasoning_role":"R1","topic":"动作预测 / 交互 World Models","url":"https://doi.org/10.1145/3757377.3764009","doi":"10.1145/3757377.3764009","generation_model":"Wan2.1 clip generator；FLUX.1-dev 背景与 Hunyuan3D 角色 assets","auxiliary_models":"Gemini-2.0-Flash director / voiceover；GPT-4o keyframes；MCTS-Gen + AniEval reviewer","conditions":"文本故事 → storyboard / visual bank / keyframes；历史 clips 与评分驱动 MCTS 选择","training_datasets":"生成 workflow 使用现成模型；AniEval 的 face-consistency evaluator 另在 Anime Face Dataset 上训练 InceptionNeXt","evaluation_datasets":"multi-shot animation、AniEval 与人类评价；故事测试集规模未核实","evidence_location":"§3.3.1–3.3.4，PDF pp.3–5","training_status":"生成器推理；evaluator 有训练","annotation_notes":"不能将整套系统笼统标为完全无训练，也不能把 Anime Face Dataset 写成 Wan 的训练集。","arxiv_id":"2506.10540","evidence_url":"https://arxiv.org/pdf/2506.10540","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"","classification_review":""},{"id":"SIGGRAPH-Asia-2025-0132","venue":"SIGGRAPH-Asia","year":2025,"title":"FairyGen: Storied Cartoon Video from a Single Child-Drawn Character","reasoning_role":"R1","topic":"通用生成框架 / 联合音视频","url":"https://doi.org/10.1145/3757377.3764007","doi":"10.1145/3757377.3764007","generation_model":"Wan-family MMDiT image-to-video + motion customization adapter；background inpainting diffusion，精确 checkpoint 未核实","auxiliary_models":"MLLM storyboard；GPT-4 风格文字描述；3D character proxy 产生 motion sequences","conditions":"单张儿童角色画 + storyboard；角色参考、背景风格与 motion customization","training_datasets":"以输入角色画学习 style adapter；由该角色 3D proxy 渲染的 motions 训练 identity / motion adapters（先打乱帧，后时序训练）；无需额外配对视频集","evaluation_datasets":"AnimatedDrawings 角色；24 style images / 12 motion videos 的比较","evidence_location":"§4–5.1，PDF pp.4–6","training_status":"per-character style / motion adapter 训练","annotation_notes":"AnimatedDrawings 是角色来源；实际 adaptation 数据是输入画及其派生渲染序列，不等同训练整个 AnimatedDrawings corpus。","arxiv_id":"2506.21272","evidence_url":"https://arxiv.org/pdf/2506.21272","evidence_level":"正文 / 附录相关段落","annotation_date":"2026-10-14","additional_evidence_url":"","unresolved_fields":"generation_model","classification_review":""}]