Image editing models have become adept at modifying visual content from natural language prompts, but dense documents remain a hard case. A single edit can touch tightly coupled text, tables, and layout, so models must make precise local changes without disrupting the surrounding page. A new benchmark, VDE Bench, is designed specifically for this setting.
The resource contains 3,506 instruction-guided editing samples spanning Chinese and English document images, split into 1,578 text-region edits and 1,928 table-region edits. Each sample pairs an editing instruction and reference image with annotated editing regions and structured OCR outputs, allowing evaluators to pinpoint exactly where a model succeeded or failed.
The evaluation protocol combines region-level and full-page measurements to cover textual correctness, structural fidelity, and visual content preservation. By focusing on the interaction between local edits and their surrounding document context, VDE Bench complements existing visual text editing benchmarks and offers a diagnostic tool for document-specific editing failures.