VAST 1.0: A Unified Framework for Controllable and Consistent Video Generation

TL;DR

The paper tackles the challenge of producing high-quality, temporally coherent video from text descriptions with precise motion control. It introduces VAST, a two-stage framework that first generates StoryForge storyboards from text and then VisionForge synthesizes video from those storyboards using diffusion-based models, thereby decoupling text understanding from video generation. Key contributions include a novel storyboard representation with explicit spatial and temporal controls, a large-scale training dataset (100 million images and 30 million videos), and state-of-the-art results on the VBench benchmark with high quality and semantic fidelity and strong temporal stability. The approach demonstrates significant improvements in dynamic control and coherence, with practical implications for VR, content creation, and simulation environments.

Abstract

Generating high-quality videos from textual descriptions poses challenges in maintaining temporal coherence and control over subject motion. We propose VAST (Video As Storyboard from Text), a two-stage framework to address these challenges and enable high-quality video generation. In the first stage, StoryForge transforms textual descriptions into detailed storyboards, capturing human poses and object layouts to represent the structural essence of the scene. In the second stage, VisionForge generates videos from these storyboards, producing high-quality videos with smooth motion, temporal consistency, and spatial coherence. By decoupling text understanding from video generation, VAST enables precise control over subject dynamics and scene composition. Experiments on the VBench benchmark demonstrate that VAST outperforms existing methods in both visual quality and semantic expression, setting a new standard for dynamic and coherent video generation.

Figures (6)

• Figure 1: Overview of the VAST Framework. The storyboard serves as an intermediary representation, providing precise control signals that enhance temporal consistency, motion dynamics, and spatial coherence in the generated videos.
• Figure 2: The StoryForge Framework for Text-to-Storyboard Generation. StoryForge converts textual descriptions into detailed storyboards, consisting of human poses, and object layouts. These intermediate representations capture the semantic and structural essence of the input scene, providing precise control signals for subsequent video generation.
• Figure 3: The VisionForge Framework for Storyboard-to-Video Generation. VisionForge takes detailed storyboards as input and synthesizes high-fidelity videos. DiT model ensures dynamic motion, temporal consistency, and spatial coherence by leveraging the structured information encapsulated in the storyboard.
• Figure 4: Qualitative Results of VAST on VBench Prompts. Examples of videos generated by VAST in response to VBench prompts, demonstrating superior temporal consistency, dynamic motion, and semantic accuracy.
• Figure 5: High-Quality Motion Generation by VAST. Videos generated by VAST demonstrating realistic motion dynamics. These examples highlight VAST's ability to intricate action sequences, capturing temporal consistency and dynamic complexity that surpasses existing methods.