ViTeX-Bench: NeurIPS-Accepted Benchmark for Video Scene Text Editing with 387 Videos

_vztu · x · 2026-10-01

• A team released ViTEX-Bench, a comprehensive benchmark for video scene text editing, accepted to NeurIPS 2026 E&D Track.
• The task is surprisingly hard: models must generate correct text, keep it temporally consistent across frames, and preserve the rest of the scene.
• The release includes:
• ViTEX-Dataset: 387 real-world 720p videos with text-region masks and editing instructions
• A three-axis evaluation protocol covering text correctness, temporal quality, and edit locality with 13 metrics
• Evaluation of 8 baselines across four video-editing paradigms
• ViTEX-Edit-14B, an open-source reference model with motion-aligned glyph-video conditioning
• Data, evaluation code, and model weights are all open-sourced.

Original post →

More from Multimodal

• Runway launches Visual Thinking: real-time video model that visualizes coding agent workflows — runwayml · 2026-10-01
• Runway demo: expand any media into an interactive real-time world simulation — runwayml · 2026-10-01
• Runway Details Portals: Draggable Windows of Generated Content With Dream-Like Edges — runwayml · 2026-10-01
• Runway Unveils Project Continuum, an OS Research App Built on Real-Time Video Interfaces — runwayml · 2026-10-01
• Qwen-Image 2.1 prompt enhancer hits 4.4x speedup in ComfyUI, now runs on 8GB VRAM — mozophe · 2026-10-01
• Seedance 2 video demo: 'An Afternoon on Skates' showcases generation quality — Ok-Nerve941 · 2026-10-01

Multimodal channel →