One Model, Three Modalities: ByteDance Releases Lance for Image and Video Understanding, Generation, and Editing

ByteDance Releases Lance, a Lightweight Native Unified Multimodal AI Model

ByteDance has released Lance, a lightweight native unified multimodal AI model with only 3 billion activated parameters, according to IT Home reporting on May 22. Unlike most existing multimodal approaches that separate "understanding" and "generation" into distinct modules and stitch them together, Lance was designed from the ground up as a unified system that handles image understanding, video understanding, image generation, video generation,…

One Model, Three Modalities: ByteDance Releases Lance for Image and Video Understanding, Generation, and Editing

Building a single model that can both understand and generate images and videos is harder than it sounds. The two tasks pull in opposite directions. Understanding benefits from high-level semantic features tightly aligned with language. Generation needs low-level continuous representations that preserve texture, geometry, and temporal dynamics. Most systems handle this tension by separating the two into distinct architectures, then bridging them…