西湖大学发布世界模型WorldForge，让普通视频模型秒变「世界引擎」

Core Viewpoint - The article discusses the advancements in AI video generation, particularly focusing on the World Forge framework developed by the West Lake University AGI Lab, which allows for precise control over video generation without sacrificing quality or requiring retraining of models [2][3][32]. Summary by Sections Introduction to AI Video Generation - Since the introduction of Sora, the realism of AI-generated videos has significantly improved, but controllability remains a challenge [2]. - Current methods either require expensive fine-tuning or lead to quality degradation due to noise and artifacts in guiding signals [2]. World Forge Framework - World Forge is a new framework that enables precise control during the video generation process without modifying model weights, effectively adding a "director's brain" to video diffusion models [3][32]. - The framework allows for the generation of 360° videos from a single image and the ability to reframe videos with complex camera movements [6][21]. Method Overview - The framework operates on a training-free guidance principle, injecting "spatiotemporal geometry" during inference [12]. - It employs a series of innovative guiding modules to ensure that the model adheres to spatial and temporal consistency while maintaining creative freedom [13]. Key Innovations 1. Intra-step Recursive Refinement (IRR): This mechanism ensures that AI-generated movements strictly follow predefined camera trajectories by incrementally correcting predictions with real content [15]. 2. Flow-Gated Latent Fusion (FLF): This module separates motion and appearance channels in the latent space, allowing precise control signals to be sent only to motion channels, preserving detail in appearance channels [16]. 3. Dual-Path Self-Correction Guidance (DSG): This strategy balances trajectory accuracy and image quality by dynamically adjusting the guiding signals based on the differences between guided and non-guided paths [17]. Performance Highlights - World Forge excels in generating 360° panoramic views from a single image, overcoming limitations of traditional panorama methods [21]. - It allows for cinematic-level video reframing, enabling users to specify complex camera movements while maintaining stability and reducing artifacts [23]. - The framework supports video editing capabilities, such as stabilizing footage, removing unwanted objects, and seamlessly integrating new elements [29]. Advantages of World Forge - The training-free nature of World Forge significantly lowers the barrier to creating high-quality 3D/4D visual content, making it accessible for various applications in film, gaming, and digital twin technologies [32][34]. - Its flexibility allows it to be integrated into various mainstream video models without the need for targeted retraining, showcasing strong generalization capabilities across different domains [34].