VFDR-Bench

VFDR-Bench (Visually Faithful Document Reconstruction Benchmark) evaluates how well a document-to-HTML system preserves a document's text, logical structure, physical layout, and visual styling when converting a page image into HTML — a representation the benchmark calls Fid-HTML.

It was introduced alongside REPLICA, an agentic framework for visually faithful document reconstruction, published at ICDAR 2026 (Oral).

• 📄 Paper: REPLICA: An Agentic Framework for Visually Faithful Document Reconstruction
• 🌐 Project page: replica-agents.github.io

This split provides a curated sample of documents and their Fid-HTML annotations for browsing and quick experimentation; the source dataset identity of each sample is intentionally not included.

What is Fid-HTML?

Fid-HTML is a semantically enriched, layout-aware HTML representation of a document page. Unlike plain OCR-to-text or OCR-to-Markdown output, it jointly encodes:

1. Textual content — full text retained at word and block level
2. Hierarchical structure — nested <div> containers reflecting sections, subsections, and logical groupings
3. Semantic tags — structured elements (<li>, <table>, <tr>, <td>, ...) with meaningful class names
4. Positional information — global placement via absolute positioning, local alignment via relative containers
5. Styling metadata — font size, color, and text attributes (bold, italic, underline, strikethrough) via inline CSS
6. Figures and backgrounds — page-level and inline images, laid into the DOM with accessibility-oriented alt text

Dataset columns

Column	Type	Description
image	image	The original document page image
fid-html	string	The Fid-HTML reconstruction of that page. Image references are placeholders (image.jpg for the page background, figure_N.jpg for inline figures with alt text) rather than embedded pixel data
fid-html-rendered-image	image	A rendering of the (unmodified) Fid-HTML, for visually sanity-checking the reconstruction against image

Benchmark coverage (full VFDR-Bench)

The full VFDR-Bench corpus spans:

• 5,000 documents
• 22 languages
• 17 domains, including scanned and born-digital sources, single- and multi-column layouts
• Fully expert human-annotated ground truth for structural, spatial, and visual correctness

This Hugging Face release is a subset for exploration and Data Studio preview; the full benchmark release is planned.

Evaluation metrics

VFDR-Bench scores a system along four axes, plus an aggregate Overall Score (OS):

Axis	Metric(s)	What it measures
Text Extraction (T.E.)	Word Recognition Rate (WRR), Character Recognition Rate (CRR)	Text transcription accuracy
Logical Structure (L.S.)	Normalized Tree Edit Distance (NTED), TEDS (tables), CDM (formulas)	Reading order, hierarchy, tables/lists/formulas
Physical Structure (P.S.)	Global Position Score (GPS), Local Position Score (LPS)	Layout- and line-level spatial fidelity (F1 of area-normalized overlap)
Visual Fidelity (V.F.)	Visual Fidelity Score (VFS), via VLM-as-a-judge	Rendered visual similarity to the ground truth

Overall = (TE + LS + PS + VF) / 4,   TE = (CRR + WRR) / 2,   PS = (GPS + LPS) / 2

Overall scores are also reported per-language subset: OS_all (all languages), OS_en (English), OS_zh (Chinese).

Benchmark results

Quantitative comparison of document-to-HTML methods on VFDR-Bench (from the paper, Table 2). Higher is better for all metrics except NTED (lower is better).

Pipeline tools

Method	WRR↑	CRR↑	NTED↓	LPS↑	GPS↑	VFS↑	OS_all↑	OS_en↑	OS_zh↑
Marker	0.68	0.82	0.61	0.19	0.34	0.74	0.53	0.60	0.45
Docling	0.40	0.49	0.62	0.17	0.26	0.55	0.38	0.45	0.12

Expert (OCR-specialist) VLMs

Method	WRR↑	CRR↑	NTED↓	LPS↑	GPS↑	VFS↑	OS_all↑	OS_en↑	OS_zh↑
rolmOCR	0.54	0.67	0.95	0.15	0.07	0.56	0.34	0.42	0.35
smolDocling	0.37	0.54	0.89	0.16	0.23	0.54	0.28	0.35	0.02
olmOCR	0.49	0.64	0.93	0.21	0.09	0.33	0.31	0.38	0.34
Nanonets-OCR-s	0.25	0.42	0.94	0.18	0.10	0.56	0.26	0.34	0.31
GOT-OCR-2.0	0.36	0.39	0.93	0.15	0.25	0.70	0.32	0.49	0.10
ChatDoc-OCRFlux	0.30	0.38	0.88	0.15	0.24	0.44	0.27	0.39	0.34

General-purpose VLMs

Method	WRR↑	CRR↑	NTED↓	LPS↑	GPS↑	VFS↑	OS_all↑	OS_en↑	OS_zh↑
Qwen3-VL	0.66	0.79	0.59	0.22	0.30	0.43	0.47	0.48	0.31
Gemma3	0.58	0.59	0.85	0.19	0.25	0.38	0.36	0.42	0.37
InternVL3	0.06	0.30	0.82	0.19	0.29	0.31	0.22	0.29	0.27
Gemini-2.5-Pro	0.67	0.42	0.56	0.14	0.31	0.82	0.47	0.47	0.30
GPT-5	0.65	0.79	0.49	0.21	0.43	0.86	0.58	0.62	0.50

Screenshot-to-HTML methods

Method	WRR↑	CRR↑	NTED↓	LPS↑	GPS↑	VFS↑	OS_all↑	OS_en↑	OS_zh↑
Webcode2m	0.33	0.45	0.82	0.15	0.29	0.39	0.30	0.42	0.13
Waffle VLM Websight	0.14	0.42	0.84	0.11	0.29	0.19	0.22	0.27	0.11

REPLICA (ours)

Method	WRR↑	CRR↑	NTED↓	LPS↑	GPS↑	VFS↑	OS_all↑	OS_en↑	OS_zh↑
REPLICA	0.88	0.93	0.20	0.73	0.86	0.93	0.86	0.93	0.89

REPLICA achieves the best overall result across all methods — near-perfect text extraction, strong logical structure preservation, physical structure recovery over 85%, and the highest visual fidelity (93%) — outperforming strong baselines including GPT-5 and Gemini-2.5-Pro. Notably, REPLICA with Qwen3-VL-8B achieves a 39% overall gain over raw Qwen3-VL-30B despite being 3.75× smaller.

Citation

@InProceedings{Raghuveer_2027_REPLICA,
    author    = {Raghuveer, R. and Srinivasan, Anirudh and Venna, Venkata Kesav and Sreevatsa, S. and Jain, Aryan and Kukkala, Sahithi and Sarvadevabhatla, Ravi Kiran},
    editor    = {Fink, Gernot A. and Forn{\'e}s, Alicia and Kise, Koichi and Lopresti, Daniel},
    title     = {REPLICA: An Agentic Framework for Visually Faithful Document Reconstruction},
    booktitle = {Document Analysis and Recognition -- ICDAR 2026},
    year      = {2027},
    publisher = {Springer Nature Switzerland},
    address   = {Cham},
    pages     = {505--522},
    isbn      = {978-3-032-36023-6},
    doi       = {10.1007/978-3-032-36023-6_29},
    url       = {https://doi.org/10.1007/978-3-032-36023-6_29}
}

License

Apache 2.0