OmniDoc-TokenBench: A Benchmark for VAE Reconstruction on Text-Rich Document Images

A curated benchmark containing approximately 3,000 samples designed to evaluate VAE reconstruction on text-rich document images. The dataset spans nine categories including books, slides, and academic papers in both English and Chinese. It was created by alibabagroup and includes an evaluation toolkit supporting metrics like PSNR, SSIM, LPIPS, and FID.

Use Cases

• Benchmarking VAE reconstruction quality based on the described metrics like PSNR and SSIM
• Training models for document image generation based on the nine described categories
• Evaluating model performance on multilingual text-rich images based on the English and Chinese content

Strengths

• Contains approximately 3,000 samples across nine distinct document categories
• Includes both English and Chinese language content, supporting multilingual evaluation
• Provides an evaluation toolkit supporting multiple established metrics like PSNR, SSIM, LPIPS, and FID

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

alibabagroup

Freshness

Last updated 2026-05-14 03:16:55