ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing
ViTeX-Bench tests whether video editors can change scene text without breaking motion, image quality, or nearby content.
The benchmark includes 387 real-world 720p videos with masks and edit instructions, split into 230 paired training examples and 157 frozen evaluation videos. Its protocol scores text correctness, visual and temporal quality, and edit locality across 13 metrics. Eight baseline systems struggled to get accurate text, stable video, and scene preservation at the same time. The authors also release ViTeX-Edit-14B, a reference editor fine-tuned on the training split. ArXiv · AI/CL/LG's note
The benchmark includes 387 real-world 720p videos with masks and edit instructions, split into 230 paired training examples and 157 frozen evaluation videos. Its protocol scores text correctness, visual and temporal quality, and edit locality across 13 metrics. Eight baseline systems struggled to get accurate text, stable video, and scene preservation at the same time. The authors also release ViTeX-Edit-14B, a reference editor fine-tuned on the training split. ArXiv · AI/CL/LG's note
score 4