AI Native Daily Paper Digest – 20250121
1. GameFactory: Creating New Games with Generative Interactive Videos
🔑 Keywords: Generative game engines, scene generalization, video diffusion models, action-controllable
💡 Category: Generative Models
🌟 Research Objective:
– To explore scene generalization in game video generation and enable the creation of diverse and interactive game content.
🛠️ Research Methods:
– Utilizes pre-trained video diffusion models and proposes a multi-phase training strategy to bridge domain gaps and achieve action controllability.
💬 Research Conclusions:
– GameFactory effectively generates diverse, open-domain, and action-controllable game videos, advancing AI-driven game creation.
👉 Paper link: https://huggingface.co/papers/2501.08325
2. VideoWorld: Exploring Knowledge Learning from Unlabeled Videos
🔑 Keywords: deep generative model, visual input, VideoWorld, Latent Dynamics Model
💡 Category: Generative Models
🌟 Research Objective:
– The study aims to explore whether a deep generative model can learn complex knowledge solely from visual input rather than text-based models.
🛠️ Research Methods:
– Developed VideoWorld, an auto-regressive video generation model trained on unlabeled video data; introduced the Latent Dynamics Model as a component to enhance knowledge acquisition from visual data.
💬 Research Conclusions:
– VideoWorld demonstrates that video-only training offers sufficient information for learning knowledge, including rules and reasoning, without search algorithms or reward mechanisms typical in reinforcement learning; achieves impressive levels in video-based tasks and generalizes effectively in robotic environments.
👉 Paper link: https://huggingface.co/papers/2501.09781
3. SEAL: Entangled White-box Watermarks on Low-Rank Adaptation
🔑 Keywords: LoRA, SEAL, Watermarking, Copyright Protection
💡 Category: AI Systems and Tools
🌟 Research Objective:
– Propose SEAL, a watermarking technique for protecting LoRA weights, addressing copyright issues.
🛠️ Research Methods:
– Embeds a secret, non-trainable matrix as a “passport” in LoRA weights.
– Entangles the passport through training without performance loss.
💬 Research Conclusions:
– SEAL maintains performance across various tasks.
– Demonstrates robustness against removal, obfuscation, and ambiguity attacks.
👉 Paper link: https://huggingface.co/papers/2501.09284
