创意人工智能研究组 AI4C Research Team

» 探索机器创造力的边界 —— 从“理解人类创作”出发，赋予AI顶层设计与美学推理的能力。 Exploring the boundaries of machine creativity — starting with 'understanding human creation' to endow AI with top-level design and aesthetic reasoning capabilities.

我们旨在打破传统AI“只能机械生成、无法理解创作”的局限。本方向不仅研究人类在艺术创作中的意图、过程与认知机制，更致力于让AI真正具备类似人类导演或设计师的“设计思维 (Design Thinking)”。 1. 跨模态美学对齐：优秀的艺术作品源于对声音、画面与故事的整体想象。我们的研究核心是实现文字、音乐、色彩、情感与动态效果等跨模态设计元素的深度对齐，赋予AI顶层的美学感知和全景式构思能力，使其不仅能生成素材，更能理解创作背后的美学逻辑。 2. 大模型的可解释性与轻量化： 从微观（神经元激活表征）、中观（子网络与注意力机制协同）到宏观（整体泛化与涌现能力）三个尺度，深度剖析生成式大模型的内在机理。通过揭示这些“黑盒”规律，指导算法优化与架构精简，实现“以小博大”的高效、轻量化计算。 We aim to break through the traditional limitation of AI being confined to mechanical generation without an understanding of the creative process. This direction not only investigates human intent, processes, and cognitive mechanisms in artistic creation but also strives to grant AI a 'Design Thinking' capability akin to that of human directors or designers. 1. Cross-modal Aesthetic Alignment: Outstanding artworks stem from a holistic imagination of sound, visuals, and narrative. Our core research focuses on the deep alignment of cross-modal design elements—such as text, music, color, emotion, and motion—endowing AI with top-level aesthetic perception and panoramic conceptualization abilities. This enables AI to generate assets while truly understanding the underlying aesthetic logic of creation. 2. Interpretability and Lightweighting of Large Models: We conduct a multi-scale analysis of generative large models—from the micro level (neuron activation patterns), through the meso level (sub-network and attention mechanism synergy), to the macro level (overall generalization and emergent capabilities). By demystifying these 'black boxes,' we guide algorithmic optimization and architectural refinement to achieve efficient, lightweight computing that 'does more with less.'

• 多模态感知 Multimodal Perception
• 跨模态对齐 Cross-modal Alignment
• 创作推理 Creative Reasoning
• 设计思维 Design Thinking
• 美学感知 Aesthetic Perception

» 探索人类创造力的边界 —— 从“数字资产高效生产”出发，将创意智能体融入工业流水线，解放人类生产力。 Exploring the boundaries of human creativity — starting with 'efficient digital asset production' to integrate creative agents into industrial pipelines and liberate human productivity.

我们专注于将前沿的AI能力转化为高效的生产力工具。通过解决2D/3D动画及影视制作中耗时费力的机械性劳动，让创作者从繁琐的重复性工作中抽离出来，专注于灵感本身。 1. 全维度多模态交互： 我们的研究横跨一维音频、二维原画、三维几何资产到四维时空视频。通过打破单一的“图文驱动 (Text/Image-to-X)”局限，拓展更符合艺术家直觉的交互方式（如动作、音频或多模态协同驱动）。 2. 生产应用范式的革命： 推动生成式AI与现有成熟工业流流水线的深度融合，解决视觉预览、冷启动3D建模、手工补间帧等具体痛点，实现从繁琐的“工作流 (Workflow)”到纯粹的“创意流 (Creative Flow)”的范式转变。 We focus on translating cutting-edge AI capabilities into efficient productivity tools. By addressing the labor-intensive, mechanical tasks in 2D/3D animation and film production, we enable creators to detach from tedious repetitive work and refocus on inspiration itself. 1. Full-dimensional Multimodal Interaction: Our research spans 1D audio, 2D concept art, 3D geometric assets, and 4D spatiotemporal video. Moving beyond single-modality 'Text/Image-to-X' paradigms, we explore interaction methods that align with artists' intuition, such as motion-driven, audio-driven, or multimodal collaborative generation. 2. Revolutionizing Production Paradigms: We promote the deep integration of generative AI with established industrial pipelines. By solving specific pain points—such as visual previews, cold-start 3D modeling, and manual in-betweening—we facilitate a paradigm shift from cumbersome 'workflows' to pure 'creative flow.'

• 全维度多模态交互 Full-dimensional Multimodal Interaction
• 2D/3D/4D 内容 2D/3D/4D Content
• 数字资产生产 Digital Asset Production
• 影视动画工作流 Film & Animation Workflow
• 工业流水线融合 Industrial Pipeline Integration

» 探索生命认知的边界 —— 借鉴神经生物学与认知科学，开展“类脑研究”与“科学赋能”的双向互哺。 Exploring the boundaries of life cognition — leveraging neurobiology and cognitive science to foster a bidirectional symbiosis between 'brain-inspired research' and 'scientific empowerment.'

我们聚焦于人工智能与神经科学的交叉前沿。通过将计算科学与生物智能深度结合，寻找下一代AI架构的突破口，同时利用AI技术拓展人类对大脑的认知边界。 1. 向内借鉴 (仿真驱动的类脑研究)： 引入前沿神经科学理论，将复杂的生物学神经机制转化为可计算的数学模型。通过构建对应大脑不同功能区及脑网络的数学模型与神经网络算法，寻求AI模型架构层面的突破。 2. 向外赋能 (AI驱动的神经科学)： 利用先进的生成式AI与数据分析技术，反哺神经科学研究。辅助科学家探索人类大脑复杂的神经连接机制和疾病机理，在脑科学领域发掘并解决具有重大价值的科学难题，并推动临床应用。 We focus on the intersection of artificial intelligence and neuroscience. By deeply integrating computational science with biological intelligence, we seek breakthroughs for next-generation AI architectures while utilizing AI technology to expand the boundaries of human understanding of the brain. 1. Inward Inspiration (Simulation-driven Brain-inspired Research): Introducing advanced neuroscience theories, we translate complex biological neural mechanisms into computable mathematical models. By constructing mathematical models and neural network algorithms corresponding to different functional areas and brain networks, we seek architectural breakthroughs in AI models. 2. Outward Empowerment (AI-driven Neuroscience): Utilizing advanced generative AI and data analytics, we give back to neuroscience research. We assist scientists in exploring the complex neural connectivity mechanisms and disease pathologies of the human brain, aiming to solve significant scientific challenges in brain science and promote clinical applications.

• 类脑研究 Brain-inspired Research
• 神经科学 Neuroscience
• 双向互哺 Bidirectional Symbiosis