A video editor that replaces text in moving footage must get the new wording right while preserving the scene. A new benchmark from Texas A&M researchers shows that partial character similarity and exact-text matches can rank systems differently, though the reported differences do not establish a clear winner.
ViTeX-Bench evaluates 387 real-world 720p clips, with text regions marked for editing and instructions specifying the original and replacement wording. Of those clips, 230 reviewed examples with paired edits form the training set; the other 157 are held out for testing. The benchmark scores 13 measures of text accuracy, temporal quality and how much of the surrounding scene changes, comparing eight existing methods with the researchers’ ViTeX-Edit-14B reference model.
On the text measures, ViTeX-Edit-14B had higher mean character similarity than VideoPainter (0.688 versus 0.619), but a lower score for reproducing the exact requested string (0.341 versus 0.364). Text-reading software detected no original wording in five test clips, so the clips were excluded from both text measures. The paper reports overlapping 95% confidence intervals, so these scores do not establish a statistically significant winner. Partial character similarity is not the same as getting the full requested wording right.
The reference editor uses a sequence of replacement-letter shapes aligned with the text across frames to guide the edit. The team released its model weights, code and dataset. The benchmark mainly covers localized, readable text in Latin scripts; dense layouts, curved surfaces, heavy occlusion, extreme motion and other writing systems are sparsely represented.
