AI Open‑Source Forum Recap: Video Generation, Vision, Vector DBs, AI‑Native Language

Amid an unprecedented AI boom driven by large‑scale models, multimodal foundations, and AI‑native software stacks, the industry and academia are converging, making research‑industry collaboration and open‑source ecosystems the most critical innovation paradigm of the era.

Shi Baixin (Peking University) highlighted the limitations of current video generation—spatial distortion, discontinuous view stitching, inconsistent motion, and poor long‑term stability. His team introduced an interval‑flow technique that lets the model attend to multiple frames before and after, establishing cross‑time attention. They built a multi‑stream temporal control network that processes demixed audio tracks with self‑attention, enabling precise lip‑sync, event sequencing, and emotional rendering. A new dataset of ~392,000 clips (~1,200 h) was created, supporting facial sync, temporal control, and emotion rendering. Their work, titled “Audio‑Sync Video Generation with Multi‑Stream Temporal Control,” was accepted to NeurIPS 2025. Additionally, they demonstrated panoramic video generation with latitude/longitude‑aware mechanisms, also accepted to NeurIPS 2025.

Li Chenglong (Zilliz) recapped Milvus, the world’s first open‑source vector database launched in October 2019. Milvus now serves over 10,000 enterprises and has amassed 40 K GitHub stars. It supports Float, Sparse, Binary vectors, dynamic insert/delete, tag‑plus‑vector filtering, and keyword‑plus‑vector search. The 2021 LTS release added data persistence, sharding, and heterogeneous hardware support, but its all‑in‑one architecture limited scalability for large‑scale or high‑QPS workloads. In Milvus 2.6, the team added a StreamingNode for incremental data, merged DataNode and IndexNode, and introduced a custom Woodpecker message queue. Zilliz has commercialized the project by offering Milvus as a fully managed SaaS on public clouds (Zilliz Cloud), turning open‑source adoption into enterprise SaaS revenue.

Chen Hui (Tsinghua University) addressed the urgent need for lightweight, high‑performance vision backbones for edge devices. His team proposed an asymmetric deep learning structure that uses a complex architecture during training and an equivalent compressed network for inference, reducing redundancy. They released RepViT (CVPR 2024) and LSNet (CVPR 2025). For object detection, they tackled YOLO’s multi‑box fusion and NMS dependency by introducing a consistent double‑label matching strategy, enabling an NMS‑free YOLOv10 (NeurIPS 2024) that balances state‑of‑the‑art accuracy with inference speed. They also built YOLOE (ICCV 2025), an open‑domain visual understanding model that leverages large language models for multimodal prompts, supporting both detection and segmentation in open scenes.

Lei Zhengyu (MoonBit) explained how AI is reshaping software development from “human‑write‑code + machine assistance” to “AI‑generated, AI‑reviewed” workflows. MoonBit aims to create an AI‑native programming language with ultra‑fast compilation, small binary output, built‑in static analysis, a gentle learning curve, and expressive power without relying on conventions. The community has produced thousands of open‑source packages for web development, numerical computing, and middleware SDKs, and integrates with Python, JavaScript, and WebAssembly to reduce cross‑language duplication and compatibility costs.

The forum concluded with a promise to release full video recordings of the talks, providing deeper insight into the presented research and open‑source projects.