Research Article
Performance Comparison of Text-to-Video Generative Models
@INPROCEEDINGS{10.4108/eai.19-12-2025.2362706, author={Gourav Kashiv and Sarbjeet Singh}, title={Performance Comparison of Text-to-Video Generative Models}, proceedings={Proceedings of the 1st International Conference on Computational Techniques in Data Science, IICTDS 2025, 19-20 December 2025, Chandigarh, India}, publisher={EAI}, proceedings_a={IICTDS}, year={2026}, month={11}, keywords={Diffusion models low-rank adaptation large language models feature alignment}, doi={10.4108/eai.19-12-2025.2362706} }- Gourav Kashiv
Sarbjeet Singh
Year: 2026
Performance Comparison of Text-to-Video Generative Models
IICTDS
EAI
DOI: 10.4108/eai.19-12-2025.2362706
Abstract
Text-to-video (T2V) generation has been significantly improved using diffusion-based generative models, but the generation of high-quality videos remains a difficult challenge. This paper examines how LoRA-FA can be used to fine-tune multiple T2V generative models, such as ModelScope, Zeroscope-320, and Zeroscope-576 with fewer parameters and under limited resources. Through quantitative and qualitative analysis we demonstrate that by applying Lora-FA, semantic alignment and perceptual sharpness can be improved, effectively increasing the quality of video frames generated. We found out that the Zeroscope v2(576w) model produced significant improvements in semantic alignment and perceptual sharpness while the improvements made by ModelScope were more balanced. The findings also indicate that LoRA-FA is useful for enhancing the quality of videos in low-resource configurations without having to retrain a whole system and lead the way to further studies of multimodal, domain-specific video generation tasks, which offers an efficient, scalable structure for effective T2V model optimization.