Shilong Zhang (张士龙)

My research agenda connects visual perception/understanding and generation, aiming to build unified models that truly understand the world, rather than merely translating between modalities.

• 2020–2022 · Object Detection
  My research began with object detection, focusing on efficient architectures and training.
• 2022–2023 · Multimodal Understanding
  With the rise of large language models around 2022, I recognized that multimodal large language models would be central to the future of AI, which led me to explore vision–language models. However, I soon realized a fundamental limitation of dominant VLM paradigms: projecting vision into language space does not genuinely enable models to learn visual knowledge.
• 2023–2025 · Generative Models
  This insight motivated my transition to vision generation, where models are forced to internalize visual structure, semantics, and dynamics directly from data—much closer to how humans learn by continuously predicting visual content.
• 2025 · Unify
  After accumulating substantial experience in generative modeling, by 2025 it became clear—to both the community and myself—that understanding and generation should be unified. I therefore focus on what I consider the most urgent problem in unified modeling: a single, principled visual encoder that supports both understanding and generation. This led to PS-VAE, a key step toward a unified encoder enabling understanding, generation, and editing within a shared representation space.
• Next · Video Unified Models
  Looking forward, I believe images alone are insufficient for learning rich and grounded world knowledge, and I am particularly excited about video-based unified models, which I see as a potential next major leap beyond LLMs by capturing dynamics, causality, and long-term structure.

If your team shares this perspective, I would be glad to connect and discuss potential collaborations.