Uni-ViGU: Towards Unified Video Generation and Understanding via A Diffusion-Based Video Generator

Recent advances in video generation and video understanding have largely been driven by separate model families — diffusion-based generators and autoregressive language models, respectively. In this work, we present Uni-ViGU, a unified framework that bridges video generation and understanding within a single diffusion-based video generator. Our key insight is that a diffusion model, traditionally used only for pixel-level generation, can be extended to jointly produce both visual and textual outputs from a shared prompt, enabling a seamless integration of generation and understanding capabilities.

Through a carefully designed architecture and training strategy, Uni-ViGU generates high-quality videos while simultaneously producing coherent textual descriptions — all within one forward diffusion/denoising process. The model leverages flow matching to iteratively refine both modalities, and we demonstrate that video and text co-evolve meaningfully across denoising steps. Extensive experiments show that Uni-ViGU achieves competitive results on both video generation and understanding benchmarks, offering a new paradigm for multimodal unification.