我们介绍了[Cosmos-Predict2.5]，这是用于物理 AI 的 Cosmos 世界基础模型的最新版本。基于流式架构，[Cosmos-Predict2.5]在单个模型中统一了 Text2World、Image2World 和 Video2World 生成，并利用[Cosmos-Reason1]（一个物理 AI 视觉语言模型）来提供更丰富的文本基础和更精细的世界模拟控制。在 2 亿个精选视频片段上训练，并通过基于强化学习的后训练进行优化，[Cosmos-Predict2.5]在视频质量和指令对齐方面相比[Cosmos-Predict1]取得了显著改进，发布了 2B 和 14B 规模的模型。这些能力使得机器人技术和自主系统能够生成更可靠的合成数据、进行策略评估和闭环模拟。我们进一步扩展了该系列，推出了[Cosmos-Transfer2.5]，这是一个用于 Sim2Real 和 Real2Real 世界翻译的控制网络风格框架。尽管比[Cosmos-Transfer1]小 3.5 倍，但它提供了更高的保真度和稳健的长时域视频生成。总之，这些进展使[Cosmos-Predict2.5]和[Cosmos-Transfer2.5]成为扩展具身智能的多功能工具。为了加速物理 AI 的研究和部署，我们在 NVIDIA 开放模型许可下发布了源代码、预训练检查点和精选基准，链接为 https://github.com/nvidia-cosmos/cosmos-predict2.5 和 https://github.com/nvidia-cosmos/cosmos-transfer2.5。我们希望这些开放资源能够降低采用门槛，并促进构建下一代具身智能的创新。
We introduce [Cosmos-Predict2.5], the latest generation of the Cosmos World Foundation Models for Physical AI. Built on a flow-based architecture, [Cosmos-Predict2.5] unifies Text2World, Image2World, and Video2World generation in a single model and leverages [Cosmos-Reason1], a Physical AI vision-language model, to provide richer text grounding and finer control of world simulation. Trained on 200M curated video clips and refined with reinforcement learning-based post-training, [Cosmos-Predict2.5] achieves substantial improvements over [Cosmos-Predict1] in video quality and instruction alignment, with models released at 2B and 14B scales. These capabilities enable more reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems. We further extend the family with [Cosmos-Transfer2.5], a control-net style framework for Sim2Real and Real2Real world translation. Despite being 3.5$\times$ smaller than [Cosmos-Transfer1], it delivers higher fidelity and robust long-horizon video generation. Together, these advances establish [Cosmos-Predict2.5] and [Cosmos-Transfer2.5] as versatile tools for scaling embodied intelligence. To accelerate research and deployment in Physical AI, we release source code, pretrained checkpoints, and curated benchmarks under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-predict2.5 and https://github.com/nvidia-cosmos/cosmos-transfer2.5. We hope these open resources lower the barrier to adoption and foster innovation in building the next generation of embodied intelligence.
核心贡献 · Key contributions
引入 Cosmos-Predict2.5，基于 flow matching 的世界基础模型，统一 Text2World、Image2World 和 Video2World 生成。 Introduces Cosmos-Predict2.5, a flow-based world foundation model unifying Text2World, Image2World, and Video2World generation.
利用 Cosmos-Reason1 视觉语言模型，实现更丰富的文本 grounding 和更精细的世界模拟控制。 Leverages Cosmos-Reason1 VLM for richer text grounding and finer control over world simulation.
使用 200M 精选视频剪辑训练，结合基于强化学习的后训练，在视频质量和指令对齐上取得显著提升。 Trained on 200M curated video clips with reinforcement learning-based post-training, achieving substantial improvements in video quality and instruction alignment.
发布 2B 和 14B 参数规模的模型，支持机器人学和自主系统的可靠合成数据生成、策略评估和闭环仿真。 Releases models at 2B and 14B scales, enabling reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems.
扩展 Cosmos-Transfer2.5，一种 ControlNet 风格的 Sim2Real 和 Real2Real 转换框架，体积缩小 3.5 倍但保真度更高。 Extends with Cosmos-Transfer2.5, a ControlNet-style framework for Sim2Real and Real2Real translation, 3.5x smaller but higher fidelity.
在 NVIDIA 开放模型许可下开源代码、预训练检查点和基准，加速物理 AI 研究。 Open-sources code, pretrained checkpoints, and benchmarks under NVIDIA Open Model License to accelerate Physical AI research.
局限 · Limitations
严格的数据过滤仅保留约 4%的原始剪辑，可能降低多样性和稀有场景覆盖。 Strict data filtering retains only ~4% of raw clips, potentially reducing diversity and coverage of rare scenarios.
域特化微调和模型合并依赖超参数搜索，可能无法保证所有域上的最优性能。 Domain-specific fine-tuning and model merging rely on hyperparameter sweeps, which may not guarantee optimal performance across all domains.
通过自回归分块生成的长视频仍然存在误差累积，尽管相比先前工作有所减少。 Long video generation via autoregressive chunking still exhibits error accumulation, though reduced compared to prior work.
评估主要在 PAI-Bench 和精心设计的测试集上进行；对更广泛真实世界场景的泛化能力尚需充分验证。 Evaluation mainly on PAI-Bench and curated test sets; generalization to broader real-world scenarios remains to be fully validated.
论文章节 · Sections（共 28）
摘要Abstract
2 数据2 Data
3 方法3 Method
4 训练4 Training
6 应用6 Applications
8 结论8 Conclusion
1 引言1 Introduction
2 数据2 Data
2.1 视频整理流水线2.1 Video Curation Pipeline
2.2 领域特定数据2.2 Domain Specific Data
6 关键对象之间的交互6 Interactions between key objects.
3 方法3 Method
3.1 流匹配3.1 Flow Matching
3.2 网络架构3.2 Network Architecture
4 训练4 Training
4.1 预训练4.1 Pre-training
4.2 后训练4.2 Post-training
4.3 基础设施4.3 Infrastructure
5 结果5 Results
6 应用6 Applications
6.1 Cosmos-Transfer2.56.1 Cosmos-Transfer2.5
6.2 Cosmos-Transfer2.5 用于机器人策略学习6.2 Cosmos-Transfer2.5 for Robot Policy Learning
6.3 用于驾驶仿真的 Cosmos-Transfer2.56.3 Cosmos-Transfer2.5 for Driving Simulation
6.4 带相机控制的多视角生成6.4 Multi-view Generation with Camera Control
6.5 用于 VLA 训练的合成数据生成6.5 Synthetic Data Generation for VLA training
6.6 动作条件的世界生成6.6 Action-Conditioned World Generation