Sijie Xu
Algorithm Engineer, Super Intelligence – AIGC Group at Xiaohongshu
Algorithm Engineer at Xiaohongshu, currently the lead of the open-source project OpenStoryline. Focused on the design and deployment of conversational video-editing agents and automated creative workflows, dedicated to transforming complex video production processes into interactive intelligent systems. Has published papers at leading conferences such as ICCV and ACM Multimedia. Previously worked on AIGC-related research, including image and video stylization as well as diffusion model acceleration.
Topic
How to Build the Next-Generation Intelligent Editing System
This talk draws on real-world experience from the open-source project OpenStoryline to explore the fundamental differences between intelligent editing agents and traditional general-purpose agents. Editing systems inherently possess complete workflow structures, but they cannot be rigidly linear—each node may be skipped, rolled back, or dynamically rearranged. We will focus on how to design a “skippable” node architecture, implement dynamic scheduling and state awareness, build controllable execution loops across multi-tool, multi-stage collaboration, and establish engineering-level fallback mechanisms when models are unstable or tools fail. The talk will break down a practical intelligent editing agent design from three perspectives: system architecture, scheduling strategies, and exception recovery mechanisms. Outline: 1. Why an Intelligent Editing Agent Cannot Be a Traditional Workflow 2. Architecture Abstraction: “A Workflow, but Not a Workflow” 3. Skippable Node Design: State-Driven Rather Than Sequence-Driven 4. Dynamic Scheduling and Dependency Management Between Nodes 5. Fault Tolerance and Fallback Strategies in Multi-Stage Collaboration 6. From Experimental Demo to Production System: OpenStoryline’s Architectural Evolution 7. Reflections on the Next Generation of Creative Agents
