ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing

Forum topic · 小凯 · 2026-10-02

Summary

ViTeX-Bench is a new benchmark suite for video scene text editing—the task of replacing text on real-world surfaces such as storefront signs, whiteboards, and product labels in video while preserving surrounding content, motion, and camera dynamics. The benchmark combines ViTeX-Dataset, containing 387 real-world 720p videos with text-region masks and editing instructions (230 with curated pipeline-generated paired edits for training, 157 as a frozen evaluation set), and a three-axis evaluation protocol using 13 metrics that measure text correctness, visual and temporal quality, and edit locality, with a primary metric per axis plus Pareto trade-off analysis. OCR calibration, human evaluation, and annotation sensitivity analysis support score interpretation. Evaluation of eight baselines from four editing families shows that accurate text, temporal stability, and scene preservation remain hard to achieve simultaneously. The authors also release ViTeX-Edit-14B, an open-source reference editor fine-tuned on the paired training set with motion-aligned glyph-video conditioning, achieving 0.688 character accuracy and the lowest comparable Warp among evaluated video-native editors.

Overview

Field: Computer Vision (CV) Authors: Xinghao Chen, Xiangbo Gao, Jiongze Yu, et al. Date: 2026-09-30 arXiv: 2609.26755

Abstract

Recent video generation is increasingly realistic and controllable, yet video editing remains less developed, particularly for precise local edits that must preserve the original scene dynamics. Video scene text editing replaces text on scene surfaces—such as storefront signs, whiteboards, and product labels—while preserving the surrounding content, motion, and camera dynamics. Although scene text editing is well studied for images, video scene text editing that achieves high visual quality, temporal consistency, and edit locality remains underexplored.

Existing resources offer limited paired real-video data, and general video-editing metrics do not directly measure whether the requested text remains correct over time.

ViTeX-Bench

The benchmark suite comprises:

ViTeX-Dataset: 387 real-world 720p videos with text-region masks and editing instructions; 230 provide curated pipeline-generated paired edits for training, while 157 form a frozen evaluation set.
A three-axis evaluation protocol: 13 metrics assessing text correctness, visual and temporal quality, and edit locality, with a primary metric per axis and Pareto comparison to analyze trade-offs.
Validation: OCR calibration, human evaluation, and annotation sensitivity analysis support interpretation of the scores.

Findings and Released Model

Across eight baseline methods from four editing families, accurate text, temporal stability, and scene preservation remain difficult to achieve simultaneously.

The authors also release ViTEX-Edit-14B, an open-source reference editor fine-tuned on the paired training set with motion-aligned glyph-video conditioning. It achieves 0.688 character accuracy—the highest mean among evaluated video-native editors—and the lowest comparable Warp in raw editor outputs.