Enhancing Video Consistency in Zero-Shot Text-to-Video: Dynamic Frame Synthesis and Weighted Cross-Frame Attention over Diffusion Models.

Saved in:

Bibliographic Details
Title:	Enhancing Video Consistency in Zero-Shot Text-to-Video: Dynamic Frame Synthesis and Weighted Cross-Frame Attention over Diffusion Models.
Authors:	Wang, Shuya¹ wangsyz@njupt.edu.cn, Hu, Yunfei² hyf_njupt@163.com, Wang, Fujun³ wangfj@ypi.edu.cn
Source:	IAENG International Journal of Computer Science. Feb2026, Vol. 53 Issue 2, p576-588. 13p.
Subjects:	Stochastic models, Image quality analysis
Abstract:	Text-to-video (T2V) synthesis represents a frontier technology in visual content generation with transformative potential across advertising, gaming, education, and film production. However, current T2V methods predominantly depend on large-scale, high-quality text-video paired datasets for supervised training, imposing not only substantial computational demands but also stringent requirements on dataset scale and quality. In resource-constrained environments, zero-shot text-tovideo (ZST2V) generation offers a promising alternative, though existing approaches often exhibit limitations in temporal consistency, visual coherence, and synthesis efficiency. To overcome these challenges, this paper presents a novel diffusion-based framework for ZST2V generation. Our method decomposes video synthesis into structured image generation sub-tasks, utilizing the intrinsic noise introduction and denoising properties of diffusion models to dynamically produce frame sequences with improved content consistency. Additionally, we introduce a weighted cross-frame attention mechanism that effectively integrates features from both initial and preceding frames, enabling enhanced global context modeling and local temporal coherence. Comprehensive evaluations on the MSR-VTT and UCF101 datasets demonstrate that our approach achieves state-of-the-art performance across multiple metrics--including SSIM, PSNR, and CLIP Score--significantly advancing the visual quality, temporal stability, and semantic fidelity of zeroshot video generation. Additionally, on the UCF101 dataset, our method also achieves competitive results in terms of the IS and FVD evaluation metrics, further validating its effectiveness. [ABSTRACT FROM AUTHOR]
	Copyright of IAENG International Journal of Computer Science is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	Text-to-video (T2V) synthesis represents a frontier technology in visual content generation with transformative potential across advertising, gaming, education, and film production. However, current T2V methods predominantly depend on large-scale, high-quality text-video paired datasets for supervised training, imposing not only substantial computational demands but also stringent requirements on dataset scale and quality. In resource-constrained environments, zero-shot text-tovideo (ZST2V) generation offers a promising alternative, though existing approaches often exhibit limitations in temporal consistency, visual coherence, and synthesis efficiency. To overcome these challenges, this paper presents a novel diffusion-based framework for ZST2V generation. Our method decomposes video synthesis into structured image generation sub-tasks, utilizing the intrinsic noise introduction and denoising properties of diffusion models to dynamically produce frame sequences with improved content consistency. Additionally, we introduce a weighted cross-frame attention mechanism that effectively integrates features from both initial and preceding frames, enabling enhanced global context modeling and local temporal coherence. Comprehensive evaluations on the MSR-VTT and UCF101 datasets demonstrate that our approach achieves state-of-the-art performance across multiple metrics--including SSIM, PSNR, and CLIP Score--significantly advancing the visual quality, temporal stability, and semantic fidelity of zeroshot video generation. Additionally, on the UCF101 dataset, our method also achieves competitive results in terms of the IS and FVD evaluation metrics, further validating its effectiveness. [ABSTRACT FROM AUTHOR]
ISSN:	1819656X