diff --git a/word_html_converter/Dockerfile b/word_html_converter/Dockerfile
new file mode 100644
index 0000000..e801c48
--- /dev/null
+++ b/word_html_converter/Dockerfile
@@ -0,0 +1,13 @@
+FROM python:3.12-slim
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends libreoffice libreoffice-writer fonts-noto-cjk fonts-dejavu fontconfig \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+COPY requirements.txt /app/requirements.txt
+RUN pip install --no-cache-dir -r /app/requirements.txt
+COPY converter.py /app/converter.py
+COPY demo /app/demo
+
+ENTRYPOINT ["python", "/app/converter.py"]
diff --git a/word_html_converter/README.md b/word_html_converter/README.md
new file mode 100644
index 0000000..b336116
--- /dev/null
+++ b/word_html_converter/README.md
@@ -0,0 +1,193 @@
+# Word ↔ HTML High-Fidelity Converter
+
+一个基于开源工具的 `.docx` 与 `.html` 双向转换程序，目标是尽量保持 Word 文档在浏览器和再次转回 Word 后的视觉效果一致。
+
+> 现实说明：DOCX 和 HTML/CSS 的排版模型不同，严格意义上的“完全无损双向转换”不能只靠普通 HTML 实现。本项目采用 LibreOffice Writer 的转换引擎做高保真转换，再做 HTML/CSS 后处理；同时提供 PDF 渲染级视觉对比报告，便于验收。
+
+## 功能
+
+- DOCX → HTML
+  - 保留字体、字号、颜色、粗体、斜体、下划线、删除线
+  - 保留段落对齐、缩进、行距、页边距、分页符
+  - 保留表格结构、边框、合并单元格、部分背景色样式
+  - 保留图片、大小、位置、页眉页脚文本
+  - 将 `<style>` 中的规则尽量内联到元素 `style` 中，降低浏览器打开时样式丢失概率
+
+- HTML → DOCX
+  - 使用 LibreOffice 的 Writer HTML 导入器，解析 HTML/CSS 并导出 DOCX
+  - 对本程序生成的 HTML 回转 DOCX，视觉一致性更好
+
+- 验收校验
+  - DOCX → PDF 渲染
+  - 将原始 DOCX 与回转 DOCX 的 PDF 页面转图片后做像素差异报告
+
+## 依赖
+
+### 1. 安装 LibreOffice
+
+Linux 示例：
+
+```bash
+sudo apt-get update
+sudo apt-get install -y libreoffice libreoffice-writer fonts-noto-cjk fontconfig
+```
+
+macOS：安装 LibreOffice 后，通常路径为：
+
+```bash
+/Applications/LibreOffice.app/Contents/MacOS/soffice
+```
+
+Windows：安装 LibreOffice 后，通常路径为：
+
+```powershell
+C:\Program Files\LibreOffice\program\soffice.com
+```
+
+如程序找不到 LibreOffice，可以设置：
+
+```bash
+export SOFFICE_BIN=/path/to/soffice
+```
+
+### 2. 安装 Python 依赖
+
+```bash
+pip install -r requirements.txt
+```
+
+## 快速使用
+
+### DOCX 转 HTML
+
+```bash
+python converter.py docx2html demo/demo.docx output/demo.html
+```
+
+### HTML 转 DOCX
+
+```bash
+python converter.py html2docx output/demo2.html output/demo2.back.docx
+```
+
+### 一键往返转换
+
+```bash
+python converter.py roundtrip demo/demo.docx output/roundtrip
+```
+
+### 一键往返并生成视觉差异报告
+
+```bash
+python converter.py roundtrip demo/demo.docx output/roundtrip --verify
+```
+
+生成文件示例：
+
+```text
+output/roundtrip/demo.html
+output/roundtrip/demo.roundtrip.docx
+output/roundtrip/demo.visual-report.json
+```
+
+视觉报告字段说明：
+
+```json
+{
+  "exact_page_count": true,
+  "page_reports": [
+    {
+      "page": 1,
+      "same_size": true,
+      "mean_abs_diff_0_255": 4.2738
+    }
+  ]
+}
+```
+
+- `exact_page_count=true`：页数一致
+- `same_size=true`：渲染页面尺寸一致
+- `mean_abs_diff_0_255`：平均像素差异，越低越接近；复杂文档可设自己的验收阈值
+
+## 生成测试 DOCX
+
+项目里已经包含 `demo/demo.docx`。如果要重新生成：
+
+```bash
+python demo/make_demo_docx.py
+```
+
+## Docker 使用
+
+构建镜像：
+
+```bash
+docker build -t word-html-converter .
+```
+
+DOCX 转 HTML：
+
+```bash
+docker run --rm -v "$PWD:/work" word-html-converter \
+  docx2html /work/demo/demo.docx /work/output/demo.html
+```
+
+HTML 转 DOCX：
+
+```bash
+docker run --rm -v "$PWD:/work" word-html-converter \
+  html2docx /work/output/demo.html /work/output/demo.back.docx
+```
+
+## 关键设计说明
+
+### 为什么不用 Mammoth 或 Pandoc 作为主引擎？
+
+- Mammoth 更适合把 DOCX 转成语义清晰的 HTML，它明确不是为了逐像素复制 Word 样式。
+- Pandoc 很适合文档格式互转，但更偏结构化内容转换，不适合要求高度还原 Word 页面排版的场景。
+- LibreOffice Writer 的 DOCX/HTML 导入导出更接近真实办公软件排版结果，所以本项目把它作为主转换内核。
+
+### 为什么 HTML 转 DOCX 要强制 `HTML (StarWriter)` 输入过滤器？
+
+LibreOffice 默认可能把 HTML 当作 Web 文档打开，导致导出 DOCX 时出现“no export filter”或排版丢失。本项目使用：
+
+```bash
+--infilter="HTML (StarWriter)"
+```
+
+让 HTML 作为 Writer 文档导入，再导出 Office Open XML DOCX。
+
+### 可选归档模式
+
+DOCX 转 HTML 时可以加：
+
+```bash
+python converter.py docx2html input.docx output.html --embed-source
+```
+
+这样会把原始 DOCX 以 base64 形式嵌入 HTML。HTML 转 DOCX 时可加：
+
+```bash
+python converter.py html2docx output.html restored.docx --prefer-embedded-source
+```
+
+这适合“HTML 只用于预览/存档，希望完全恢复原始 DOCX”的场景。注意：如果用户在 HTML 中编辑了内容，使用该模式会恢复原 DOCX，不会合并 HTML 编辑内容。
+
+## 已知边界
+
+以下内容在开源转换链路中很难保证完全一致，需要单独测试：
+
+- 复杂浮动图片、环绕方式、文本框、艺术字、SmartArt
+- Word 域、目录、脚注尾注、批注、修订痕迹
+- 复杂多级编号、样式继承、主题字体
+- 页面级精确排版，如不同 Word/LibreOffice 版本的字体度量差异
+- 浏览器编辑 HTML 后再转 DOCX，不能保证所有 CSS 都能被 Writer 完整识别
+
+## 推荐验收标准
+
+建议不要用“字节级相同”验收 DOCX，因为二次生成的 DOCX 内部 XML 顺序、关系 ID、压缩结果通常会变。建议用：
+
+1. 页数一致；
+2. 关键表格行列、合并单元格、图片数量一致；
+3. PDF 渲染视觉差异低于业务阈值；
+4. 典型复杂样例人工抽检。
diff --git a/word_html_converter/converter.py b/word_html_converter/converter.py
new file mode 100755
index 0000000..0245b7b
--- /dev/null
+++ b/word_html_converter/converter.py
@@ -0,0 +1,596 @@
+#!/usr/bin/env python3
+"""
+word-html-converter
+Open-source DOCX <-> HTML conversion wrapper focused on high visual fidelity.
+
+Core engine: LibreOffice headless conversion.
+HTML post-processing: inline CSS, keep @page rules, normalize metadata.
+
+Commands:
+  python converter.py docx2html input.docx output.html
+  python converter.py html2docx input.html output.docx
+  python converter.py roundtrip input.docx output_dir
+  python converter.py verify original.docx converted.docx report.json
+"""
+from __future__ import annotations
+
+import argparse
+import base64
+import hashlib
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import warnings
+import tempfile
+from dataclasses import dataclass, asdict
+from pathlib import Path
+from typing import Dict, Iterable, List, Optional, Tuple
+
+from bs4 import BeautifulSoup, XMLParsedAsHTMLWarning
+import tinycss2
+
+
+warnings.filterwarnings("ignore", category=XMLParsedAsHTMLWarning)
+
+APP_NAME = "word-html-converter"
+DEFAULT_DOCX_TO_HTML_FILTERS = [
+    # Writer/Web HTML filter preserves more page/header/footer information than the clean XHTML filter.
+    "html",
+    # Fallback clean XHTML Writer filter.
+    "html:XHTML Writer File:UTF8",
+]
+DEFAULT_HTML_TO_DOCX_FILTERS = [
+    'docx:"Office Open XML Text"',
+    "docx",
+]
+
+
+@dataclass
+class ConvertResult:
+    source: str
+    output: str
+    command: List[str]
+    stdout: str
+    stderr: str
+    filter_name: str
+
+
+class ConversionError(RuntimeError):
+    pass
+
+
+def sha256_file(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for chunk in iter(lambda: f.read(1024 * 1024), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def find_soffice(explicit: Optional[str] = None) -> str:
+    """Find LibreOffice/soffice executable on Linux/macOS/Windows."""
+    candidates: List[str] = []
+    if explicit:
+        candidates.append(explicit)
+    if os.environ.get("SOFFICE_BIN"):
+        candidates.append(os.environ["SOFFICE_BIN"])
+
+    candidates.extend([
+        "soffice",
+        "libreoffice",
+        "/Applications/LibreOffice.app/Contents/MacOS/soffice",
+        r"C:\Program Files\LibreOffice\program\soffice.com",
+        r"C:\Program Files\LibreOffice\program\soffice.exe",
+        r"C:\Program Files (x86)\LibreOffice\program\soffice.com",
+        r"C:\Program Files (x86)\LibreOffice\program\soffice.exe",
+    ])
+    for c in candidates:
+        if not c:
+            continue
+        if Path(c).exists():
+            return str(Path(c))
+        resolved = shutil.which(c)
+        if resolved:
+            return resolved
+    raise ConversionError(
+        "LibreOffice executable not found. Install LibreOffice, or set SOFFICE_BIN=/path/to/soffice."
+    )
+
+
+def file_uri(path: Path) -> str:
+    return path.resolve().as_uri()
+
+
+def run_soffice_convert(
+    source: Path,
+    outdir: Path,
+    convert_to: str,
+    soffice_bin: Optional[str] = None,
+    timeout: int = 120,
+    input_filter: Optional[str] = None,
+) -> Tuple[subprocess.CompletedProcess[str], List[str]]:
+    source = source.resolve()
+    outdir = outdir.resolve()
+    outdir.mkdir(parents=True, exist_ok=True)
+    soffice = find_soffice(soffice_bin)
+
+    # Use an isolated LO user profile so conversion works even if desktop LO is open.
+    profile_dir = Path(tempfile.mkdtemp(prefix="lo-profile-"))
+    profile_uri = file_uri(profile_dir)
+    cmd = [
+        soffice,
+        f"-env:UserInstallation={profile_uri}",
+        "--headless",
+        "--nologo",
+        "--nofirststartwizard",
+        "--nodefault",
+        "--nolockcheck",
+        "--norestore",
+    ]
+    if input_filter:
+        cmd.append(f"--infilter={input_filter}")
+    cmd.extend([
+        "--convert-to",
+        convert_to,
+        "--outdir",
+        str(outdir),
+        str(source),
+    ])
+    try:
+        proc = subprocess.run(cmd, text=True, capture_output=True, timeout=timeout)
+    finally:
+        shutil.rmtree(profile_dir, ignore_errors=True)
+    return proc, cmd
+
+
+def find_converted_file(before: set[Path], outdir: Path, source_stem: str, expected_ext: str) -> Optional[Path]:
+    expected = outdir / f"{source_stem}{expected_ext}"
+    if expected.exists():
+        return expected
+    after = set(outdir.iterdir())
+    created = [p for p in after - before if p.is_file() and p.suffix.lower() == expected_ext.lower()]
+    if created:
+        return max(created, key=lambda p: p.stat().st_mtime)
+    candidates = list(outdir.glob(f"{source_stem}*{expected_ext}"))
+    if candidates:
+        return max(candidates, key=lambda p: p.stat().st_mtime)
+    return None
+
+
+def convert_with_fallbacks(
+    source: Path,
+    output: Path,
+    filters: Iterable[str],
+    expected_ext: str,
+    soffice_bin: Optional[str] = None,
+    timeout: int = 120,
+    input_filter: Optional[str] = None,
+) -> ConvertResult:
+    source = source.resolve()
+    output = output.resolve()
+    output.parent.mkdir(parents=True, exist_ok=True)
+
+    last_error = ""
+    with tempfile.TemporaryDirectory(prefix="convert-work-") as tmp:
+        tmpdir = Path(tmp)
+        for filter_name in filters:
+            before = set(tmpdir.iterdir())
+            proc, cmd = run_soffice_convert(source, tmpdir, filter_name, soffice_bin, timeout, input_filter=input_filter)
+            converted = find_converted_file(before, tmpdir, source.stem, expected_ext)
+            if proc.returncode == 0 and converted and converted.exists():
+                if output.exists():
+                    output.unlink()
+                shutil.move(str(converted), str(output))
+                # Move sidecar image/assets generated by HTML export.
+                for item in tmpdir.iterdir():
+                    if item.is_file() and item.name != output.name:
+                        dest = output.parent / item.name
+                        if dest.exists():
+                            dest.unlink()
+                        shutil.move(str(item), str(dest))
+                return ConvertResult(
+                    source=str(source),
+                    output=str(output),
+                    command=cmd,
+                    stdout=proc.stdout,
+                    stderr=proc.stderr,
+                    filter_name=filter_name,
+                )
+            last_error = (
+                f"Filter failed: {filter_name}\n"
+                f"Return code: {proc.returncode}\nSTDOUT:\n{proc.stdout}\nSTDERR:\n{proc.stderr}"
+            )
+    raise ConversionError(last_error or "No LibreOffice conversion filter succeeded.")
+
+
+def style_to_dict(style: str) -> Dict[str, str]:
+    result: Dict[str, str] = {}
+    if not style:
+        return result
+    declarations = tinycss2.parse_declaration_list(style, skip_comments=True, skip_whitespace=True)
+    for d in declarations:
+        if getattr(d, "type", None) == "declaration" and not d.name.startswith("--"):
+            value = tinycss2.serialize(d.value).strip()
+            if d.important:
+                value = f"{value} !important"
+            result[d.name.lower()] = value
+    return result
+
+
+def dict_to_style(d: Dict[str, str]) -> str:
+    return "; ".join(f"{k}: {v}" for k, v in d.items() if v) + (";" if d else "")
+
+
+def merge_inline_style(tag, declarations: Dict[str, str]) -> None:
+    current = style_to_dict(tag.get("style", ""))
+    current.update(declarations)
+    tag["style"] = dict_to_style(current)
+
+
+def parse_css_rules(css: str) -> Tuple[List[Tuple[str, Dict[str, str]]], str]:
+    """Return normal CSS rules plus preserved at-rules such as @page/@media."""
+    normal_rules: List[Tuple[str, Dict[str, str]]] = []
+    preserved_at_rules: List[str] = []
+    rules = tinycss2.parse_stylesheet(css, skip_comments=True, skip_whitespace=True)
+    for rule in rules:
+        if rule.type == "qualified-rule":
+            selector = tinycss2.serialize(rule.prelude).strip()
+            declarations: Dict[str, str] = {}
+            for d in tinycss2.parse_declaration_list(rule.content, skip_comments=True, skip_whitespace=True):
+                if getattr(d, "type", None) == "declaration":
+                    value = tinycss2.serialize(d.value).strip()
+                    if d.important:
+                        value = f"{value} !important"
+                    declarations[d.name.lower()] = value
+            if selector and declarations:
+                normal_rules.append((selector, declarations))
+        elif rule.type == "at-rule":
+            # @page is important for print/page-like preview; @media may contain print rules.
+            preserved_at_rules.append(tinycss2.serialize([rule]).strip())
+    return normal_rules, "\n".join(preserved_at_rules)
+
+
+def inline_css(html: str, keep_style_tag: bool = True) -> str:
+    """Inline style rules from <style> tags. This is intentionally conservative.
+
+    It handles most simple selectors generated by LibreOffice. Unsupported complex selectors
+    are skipped rather than breaking conversion.
+    """
+    soup = BeautifulSoup(html, "lxml")
+    style_tags = soup.find_all("style")
+    preserved_css: List[str] = []
+
+    for st in style_tags:
+        css = st.string or st.get_text() or ""
+        normal_rules, at_rules = parse_css_rules(css)
+        if at_rules:
+            preserved_css.append(at_rules)
+        for selector, declarations in normal_rules:
+            selectors = [s.strip() for s in selector.split(",") if s.strip()]
+            for sel in selectors:
+                # Browser-only pseudo-selectors do not make sense for DOCX reconstruction.
+                if ":" in sel and not re.search(r":(first-child|last-child|nth-child)", sel):
+                    continue
+                try:
+                    matches = soup.select(sel)
+                except Exception:
+                    continue
+                for tag in matches:
+                    merge_inline_style(tag, declarations)
+        if not keep_style_tag:
+            st.decompose()
+
+    if keep_style_tag:
+        head = soup.head or soup.new_tag("head")
+        if not soup.head and soup.html:
+            soup.html.insert(0, head)
+        if preserved_css:
+            new_style = soup.new_tag("style")
+            new_style.string = "\n".join(preserved_css)
+            head.append(new_style)
+
+    return str(soup)
+
+
+def ensure_html_metadata(html_path: Path, source_docx: Optional[Path] = None) -> None:
+    html = html_path.read_text(encoding="utf-8", errors="replace")
+    soup = BeautifulSoup(html, "lxml")
+    if soup.html is None:
+        new_html = soup.new_tag("html")
+        new_html.extend(soup.contents)
+        soup.append(new_html)
+    if soup.head is None:
+        head = soup.new_tag("head")
+        soup.html.insert(0, head)
+    if not soup.head.find("meta", attrs={"charset": True}):
+        meta = soup.new_tag("meta", charset="utf-8")
+        soup.head.insert(0, meta)
+    if not soup.head.find("meta", attrs={"name": "generator"}):
+        meta = soup.new_tag("meta")
+        meta["name"] = "generator"
+        meta["content"] = f"{APP_NAME}; LibreOffice headless"
+        soup.head.append(meta)
+    if source_docx and not soup.head.find("meta", attrs={"name": "docx-source-sha256"}):
+        meta = soup.new_tag("meta")
+        meta["name"] = "docx-source-sha256"
+        meta["content"] = sha256_file(source_docx)
+        soup.head.append(meta)
+    html_path.write_text(str(soup), encoding="utf-8")
+
+
+def embed_source_docx(html_path: Path, source_docx: Path) -> None:
+    """Optional: embed original docx for audit/archival round-trip fallback.
+
+    This does NOT merge arbitrary HTML edits back into the original docx. It is useful for
+    proving byte-exact preservation when the HTML is only used for preview/storage.
+    """
+    html = html_path.read_text(encoding="utf-8", errors="replace")
+    soup = BeautifulSoup(html, "lxml")
+    if soup.body is None:
+        body = soup.new_tag("body")
+        if soup.html:
+            soup.html.append(body)
+        else:
+            soup.append(body)
+    old = soup.find(id="__source_docx_base64__")
+    if old:
+        old.decompose()
+    script = soup.new_tag("script")
+    script["type"] = "application/vnd.openxmlformats-officedocument.wordprocessingml.document+base64"
+    script["id"] = "__source_docx_base64__"
+    script.string = base64.b64encode(source_docx.read_bytes()).decode("ascii")
+    soup.body.append(script)
+    html_path.write_text(str(soup), encoding="utf-8")
+
+
+def restore_embedded_docx_if_present(html_path: Path, output_docx: Path) -> bool:
+    html = html_path.read_text(encoding="utf-8", errors="replace")
+    soup = BeautifulSoup(html, "lxml")
+    script = soup.find(id="__source_docx_base64__")
+    if not script or not script.string:
+        return False
+    output_docx.parent.mkdir(parents=True, exist_ok=True)
+    output_docx.write_bytes(base64.b64decode(script.string.strip()))
+    return True
+
+
+def docx_to_html(
+    source_docx: Path,
+    output_html: Path,
+    soffice_bin: Optional[str] = None,
+    inline: bool = True,
+    embed_source: bool = False,
+    timeout: int = 120,
+) -> ConvertResult:
+    if source_docx.suffix.lower() != ".docx":
+        raise ConversionError("Input file must be .docx")
+    result = convert_with_fallbacks(
+        source=source_docx,
+        output=output_html,
+        filters=DEFAULT_DOCX_TO_HTML_FILTERS,
+        expected_ext=".html",
+        soffice_bin=soffice_bin,
+        timeout=timeout,
+    )
+    ensure_html_metadata(output_html, source_docx)
+    if inline:
+        html = output_html.read_text(encoding="utf-8", errors="replace")
+        output_html.write_text(inline_css(html, keep_style_tag=True), encoding="utf-8")
+    if embed_source:
+        embed_source_docx(output_html, source_docx)
+    return result
+
+
+def html_to_docx(
+    source_html: Path,
+    output_docx: Path,
+    soffice_bin: Optional[str] = None,
+    timeout: int = 120,
+    prefer_embedded_source: bool = False,
+) -> ConvertResult:
+    if source_html.suffix.lower() not in {".html", ".htm", ".xhtml"}:
+        raise ConversionError("Input file must be .html/.htm/.xhtml")
+    if prefer_embedded_source and restore_embedded_docx_if_present(source_html, output_docx):
+        return ConvertResult(
+            source=str(source_html.resolve()),
+            output=str(output_docx.resolve()),
+            command=["restore-embedded-docx"],
+            stdout="Restored embedded source DOCX.",
+            stderr="",
+            filter_name="embedded-docx",
+        )
+    return convert_with_fallbacks(
+        source=source_html,
+        output=output_docx,
+        filters=DEFAULT_HTML_TO_DOCX_FILTERS,
+        expected_ext=".docx",
+        soffice_bin=soffice_bin,
+        timeout=timeout,
+        input_filter="HTML (StarWriter)",
+    )
+
+
+def convert_to_pdf(source: Path, output_pdf: Path, soffice_bin: Optional[str] = None, timeout: int = 120) -> ConvertResult:
+    return convert_with_fallbacks(
+        source=source,
+        output=output_pdf,
+        filters=['pdf:"writer_pdf_Export"', "pdf"],
+        expected_ext=".pdf",
+        soffice_bin=soffice_bin,
+        timeout=timeout,
+    )
+
+
+def verify_docx_visual_similarity(
+    original_docx: Path,
+    converted_docx: Path,
+    report_json: Path,
+    soffice_bin: Optional[str] = None,
+    dpi: int = 120,
+    timeout: int = 120,
+) -> Dict[str, object]:
+    """Render two DOCX files to PDF, then compare page images.
+
+    Requires PyMuPDF and Pillow. The score is pragmatic, not a formal proof.
+    exact_page_count=true and low mean_abs_diff are good signs.
+    """
+    try:
+        import fitz  # type: ignore
+        from PIL import ImageChops, ImageStat  # type: ignore
+    except Exception as exc:
+        raise ConversionError("verify requires pymupdf and pillow: pip install pymupdf pillow") from exc
+
+    with tempfile.TemporaryDirectory(prefix="verify-") as tmp:
+        tmpdir = Path(tmp)
+        pdf1 = tmpdir / "original.pdf"
+        pdf2 = tmpdir / "converted.pdf"
+        convert_to_pdf(original_docx, pdf1, soffice_bin, timeout)
+        convert_to_pdf(converted_docx, pdf2, soffice_bin, timeout)
+
+        doc1 = fitz.open(str(pdf1))
+        doc2 = fitz.open(str(pdf2))
+        pages = min(len(doc1), len(doc2))
+        page_reports: List[Dict[str, object]] = []
+        zoom = dpi / 72.0
+        matrix = fitz.Matrix(zoom, zoom)
+        for i in range(pages):
+            p1 = doc1.load_page(i).get_pixmap(matrix=matrix, alpha=False)
+            p2 = doc2.load_page(i).get_pixmap(matrix=matrix, alpha=False)
+            img1 = p1.pil_image()
+            img2 = p2.pil_image()
+            same_size = img1.size == img2.size
+            if not same_size:
+                # Compare common area, record size mismatch.
+                w = min(img1.width, img2.width)
+                h = min(img1.height, img2.height)
+                img1 = img1.crop((0, 0, w, h))
+                img2 = img2.crop((0, 0, w, h))
+            diff = ImageChops.difference(img1, img2)
+            stat = ImageStat.Stat(diff)
+            mean_abs_diff = sum(stat.mean) / len(stat.mean)
+            page_reports.append({
+                "page": i + 1,
+                "same_size": same_size,
+                "mean_abs_diff_0_255": round(mean_abs_diff, 4),
+            })
+
+        report: Dict[str, object] = {
+            "original": str(original_docx.resolve()),
+            "converted": str(converted_docx.resolve()),
+            "original_sha256": sha256_file(original_docx),
+            "converted_sha256": sha256_file(converted_docx),
+            "original_pages": len(doc1),
+            "converted_pages": len(doc2),
+            "exact_page_count": len(doc1) == len(doc2),
+            "dpi": dpi,
+            "pages_compared": pages,
+            "page_reports": page_reports,
+        }
+        report_json.parent.mkdir(parents=True, exist_ok=True)
+        report_json.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+        return report
+
+
+def cmd_docx2html(args: argparse.Namespace) -> None:
+    r = docx_to_html(
+        Path(args.input),
+        Path(args.output),
+        soffice_bin=args.soffice,
+        inline=not args.no_inline_css,
+        embed_source=args.embed_source,
+        timeout=args.timeout,
+    )
+    print(json.dumps(asdict(r), ensure_ascii=False, indent=2))
+
+
+def cmd_html2docx(args: argparse.Namespace) -> None:
+    r = html_to_docx(
+        Path(args.input),
+        Path(args.output),
+        soffice_bin=args.soffice,
+        timeout=args.timeout,
+        prefer_embedded_source=args.prefer_embedded_source,
+    )
+    print(json.dumps(asdict(r), ensure_ascii=False, indent=2))
+
+
+def cmd_roundtrip(args: argparse.Namespace) -> None:
+    src = Path(args.input).resolve()
+    outdir = Path(args.output_dir).resolve()
+    outdir.mkdir(parents=True, exist_ok=True)
+    html = outdir / f"{src.stem}.html"
+    docx_back = outdir / f"{src.stem}.roundtrip.docx"
+    report = outdir / f"{src.stem}.visual-report.json"
+    r1 = docx_to_html(src, html, args.soffice, inline=not args.no_inline_css, embed_source=args.embed_source, timeout=args.timeout)
+    r2 = html_to_docx(html, docx_back, args.soffice, timeout=args.timeout, prefer_embedded_source=args.prefer_embedded_source)
+    payload = {"docx2html": asdict(r1), "html2docx": asdict(r2)}
+    if args.verify:
+        payload["verify"] = verify_docx_visual_similarity(src, docx_back, report, args.soffice, args.dpi, args.timeout)
+    print(json.dumps(payload, ensure_ascii=False, indent=2))
+
+
+def cmd_verify(args: argparse.Namespace) -> None:
+    report = verify_docx_visual_similarity(
+        Path(args.original),
+        Path(args.converted),
+        Path(args.report),
+        soffice_bin=args.soffice,
+        dpi=args.dpi,
+        timeout=args.timeout,
+    )
+    print(json.dumps(report, ensure_ascii=False, indent=2))
+
+
+def build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(description="Bidirectional high-fidelity DOCX <-> HTML converter using open-source tools.")
+    p.add_argument("--soffice", default=None, help="Path to LibreOffice soffice. Can also use SOFFICE_BIN env var.")
+    p.add_argument("--timeout", type=int, default=120, help="Conversion timeout seconds. Default: 120")
+    sub = p.add_subparsers(dest="command", required=True)
+
+    a = sub.add_parser("docx2html", help="Convert DOCX to HTML")
+    a.add_argument("input")
+    a.add_argument("output")
+    a.add_argument("--no-inline-css", action="store_true", help="Do not inline CSS from <style> tags")
+    a.add_argument("--embed-source", action="store_true", help="Embed original DOCX in HTML as base64 for archival fallback")
+    a.set_defaults(func=cmd_docx2html)
+
+    a = sub.add_parser("html2docx", help="Convert HTML to DOCX")
+    a.add_argument("input")
+    a.add_argument("output")
+    a.add_argument("--prefer-embedded-source", action="store_true", help="If HTML contains embedded original DOCX, restore it instead of converting")
+    a.set_defaults(func=cmd_html2docx)
+
+    a = sub.add_parser("roundtrip", help="DOCX -> HTML -> DOCX")
+    a.add_argument("input")
+    a.add_argument("output_dir")
+    a.add_argument("--no-inline-css", action="store_true")
+    a.add_argument("--embed-source", action="store_true")
+    a.add_argument("--prefer-embedded-source", action="store_true")
+    a.add_argument("--verify", action="store_true", help="Generate PDF-render visual diff report")
+    a.add_argument("--dpi", type=int, default=120)
+    a.set_defaults(func=cmd_roundtrip)
+
+    a = sub.add_parser("verify", help="Render two DOCX files to PDF and compare page images")
+    a.add_argument("original")
+    a.add_argument("converted")
+    a.add_argument("report")
+    a.add_argument("--dpi", type=int, default=120)
+    a.set_defaults(func=cmd_verify)
+    return p
+
+
+def main(argv: Optional[List[str]] = None) -> int:
+    args = build_parser().parse_args(argv)
+    try:
+        args.func(args)
+        return 0
+    except ConversionError as exc:
+        print(f"ERROR: {exc}", file=sys.stderr)
+        return 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/word_html_converter/demo/demo-image.png b/word_html_converter/demo/demo-image.png
new file mode 100644
index 0000000..7dce8f2
Binary files /dev/null and b/word_html_converter/demo/demo-image.png differ
diff --git a/word_html_converter/demo/demo.docx b/word_html_converter/demo/demo.docx
new file mode 100644
index 0000000..bc2791e
Binary files /dev/null and b/word_html_converter/demo/demo.docx differ
diff --git a/word_html_converter/demo/make_demo_docx.py b/word_html_converter/demo/make_demo_docx.py
new file mode 100644
index 0000000..4da447e
--- /dev/null
+++ b/word_html_converter/demo/make_demo_docx.py
@@ -0,0 +1,71 @@
+from pathlib import Path
+from docx import Document
+from docx.enum.text import WD_ALIGN_PARAGRAPH
+from docx.shared import Cm, Pt, RGBColor
+from PIL import Image, ImageDraw
+
+out_dir = Path(__file__).resolve().parent
+img_path = out_dir / "demo-image.png"
+image = Image.new("RGB", (420, 160), "white")
+d = ImageDraw.Draw(image)
+d.rectangle([10, 10, 410, 150], outline="black", width=2)
+d.text((30, 60), "DOCX embedded image", fill="black")
+image.save(img_path)
+
+doc = Document()
+section = doc.sections[0]
+section.top_margin = Cm(2)
+section.bottom_margin = Cm(2)
+section.left_margin = Cm(2)
+section.right_margin = Cm(2)
+section.header.paragraphs[0].text = "页眉：Word/HTML 双向转换测试"
+section.footer.paragraphs[0].text = "页脚：保留页眉页脚测试"
+
+style = doc.styles["Normal"]
+style.font.name = "Microsoft YaHei"
+style.font.size = Pt(12)
+
+h = doc.add_heading("Word 转 HTML 测试文件", level=1)
+h.alignment = WD_ALIGN_PARAGRAPH.CENTER
+
+p = doc.add_paragraph()
+p.paragraph_format.first_line_indent = Cm(0.74)
+p.paragraph_format.line_spacing = 1.5
+p.alignment = WD_ALIGN_PARAGRAPH.JUSTIFY
+r = p.add_run("这是一段测试文本：")
+r.font.size = Pt(12)
+r.font.color.rgb = RGBColor(0, 0, 0)
+r = p.add_run("加粗")
+r.bold = True
+r = p.add_run("、斜体")
+r.italic = True
+r = p.add_run("、下划线")
+r.underline = True
+r = p.add_run("、删除线")
+r.font.strike = True
+r = p.add_run("、红色文字。")
+r.font.color.rgb = RGBColor(192, 0, 0)
+
+for text in ["一级项目 1", "一级项目 2"]:
+    doc.add_paragraph(text, style="List Bullet")
+for text in ["编号项目 1", "编号项目 2"]:
+    doc.add_paragraph(text, style="List Number")
+
+table = doc.add_table(rows=3, cols=3)
+table.style = "Table Grid"
+for i, row in enumerate(table.rows):
+    for j, cell in enumerate(row.cells):
+        cell.text = f"R{i+1}C{j+1}"
+# Merge first row cells 1-2
+merged = table.cell(0, 0).merge(table.cell(0, 1))
+merged.text = "合并单元格"
+
+p = doc.add_paragraph("下方是一张嵌入图片：")
+p.alignment = WD_ALIGN_PARAGRAPH.LEFT
+doc.add_picture(str(img_path), width=Cm(8))
+
+doc.add_page_break()
+doc.add_paragraph("第二页内容，用于测试分页。")
+
+doc.save(out_dir / "demo.docx")
+print(out_dir / "demo.docx")
diff --git a/word_html_converter/demo/新疆玛纳斯发电有限责任公司2025年一季度经济活动分析（内部定稿）.docx b/word_html_converter/demo/新疆玛纳斯发电有限责任公司2025年一季度经济活动分析（内部定稿）.docx
new file mode 100644
index 0000000..0395225
Binary files /dev/null and b/word_html_converter/demo/新疆玛纳斯发电有限责任公司2025年一季度经济活动分析（内部定稿）.docx differ
diff --git a/word_html_converter/requirements.txt b/word_html_converter/requirements.txt
new file mode 100644
index 0000000..c87709f
--- /dev/null
+++ b/word_html_converter/requirements.txt
@@ -0,0 +1,6 @@
+beautifulsoup4>=4.12
+lxml>=5.0
+tinycss2>=1.2
+python-docx>=1.1
+pillow>=10.0
+pymupdf>=1.24