Research map / Video / Video generation
ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ViTeX-Bench, a dataset (387 real 720p videos, 230 paired training edits) and three-axis evaluation protocol for video scene text editing, plus the ViTeX-Edit-14B reference editor built on Wan2.1-VACE-14B with a glyph-video conditioning branch.
Problem
Video scene text editing (replacing text on surfaces in real videos while preserving motion, scene, and temporal consistency) lacks paired real-video data and metrics that verify whether the requested text is rendered correctly over time.
Approach
A benchmark with OCR-anchored text correctness metrics (SeqAcc, CharAcc, TTS), temporal and visual quality metrics on text crops (Flicker, Warp, MUSIQ), and edit-locality metrics on out-of-mask pixels, plus a reference editor that adds a motion-aligned glyph-video stream (homography-warped target glyphs) with a dedicated cross-attention branch on Wan2.1-VACE-14B.
Key results
- ViTeX-Edit-14B reaches CharAcc 0.688, the highest mean among video-native editors (VideoPainter 0.619).
- FLUX-Text leads SeqAcc at 0.528, with TextCtrl at 0.475; Wan2.1-VACE-14B and Kling score SeqAcc 0.
- Composite post-processing raises ViTeX-Edit-14B PSNR-loc from 29.08 to 42.95 dB while SeqAcc moves from 0.341 to 0.345.
Surprising findings
- Wan2.1-VACE-14B and Kling Video 3.0 Omni achieve SeqAcc 0 despite high visual quality, showing polish does not imply correct text.
- The Source video row (unedited) scores TTS 0.760, higher than most editors, so stability alone does not indicate editing success.
Limitations
- Coverage is predominantly Latin-script, localized, readable text; non-Latin scripts and dense or curved layouts are sparse.
- Paired training edits are model-assisted and may retain rendering errors; no ground-truth canonical renderings.
- Human locality agreement is limited (Krippendorff 0.37) and component ablations of the reference editor are not reported.
Experimental setup
Datasets: Panda-70M InternVid ViTeX-Dataset
Benchmarks: ViTeX-Bench
Metrics: SeqAcc CharAcc TTS Flicker Warp MUSIQ PSNR SSIM LPIPS DreamSim
Baselines: AnyText2 TextCtrl FLUX-Text RS-STE AnyV2V Wan2.1-VACE-14B VideoPainter Kling Video 3.0 Omni
Builds on: Wan2.1-VACE PISCO PP-OCRv5 SAM 3 STRIVE
Closest papers by meaning
- SteerVTE: Seamless Video Text Editing with Style and Glyph Control (NeurIPS 2026 · similarity 0.90)
- VE-Bench: Subjective-Aligned Benchmark Suite for Text-Driven Video Editing Quality Assessment (AAAI 2025 · similarity 0.89)
- EditBoard: Towards a Comprehensive Evaluation Benchmark for Text-Based Video Editing Models (AAAI 2025 · similarity 0.88)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.87)
- ViDiC: Video Difference Captioning (NeurIPS 2026 · similarity 0.87)
- VEU-Bench: Towards Comprehensive Understanding of Video Editing (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.