AnyDoc

Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization

Jiawei Lin^*  ·  Wanrong Zhu  ·  Vlad I. Morariu  ·  Christopher Tensmeyer

Xi'an Jiaotong University  ·  Adobe Research

_{^*Work done during an internship at Adobe Research.}

📄 Paper (CVPR 2026)  •  📄 arXiv

---

Overview

Document generation has gained growing attention in AI-driven content creation. AnyDoc is a framework capable of handling multiple generation tasks across a wide spectrum of document categories, all represented in a unified HTML/CSS format. Unlike prior work that generates documents as raster images (not editable) or flat coordinate sequences (poor at complex layouts), AnyDoc adopts a hierarchical, multi-layered HTML/CSS representation that is both expressive and fully editable.

AnyDoc contributes:

• DocHTML — a large-scale, automatically synthesized dataset of 265,206 document samples spanning 111 categories and 32 styles, each with rich metadata, HTML/CSS source code, visual assets, and rendered screenshots.
• Supervised fine-tuning (SFT) of a multi-modal LLM (Qwen2.5-VL-7B-Instruct) on DocHTML for three practical tasks.
• Height-Aware Reinforcement Learning (HARL) — a GRPO-based post-training procedure that penalizes content overflow using a reward based on the difference between predicted and target document heights.

AnyDoc outperforms both general-purpose MLLMs and task-specific baselines across all three tasks.

Tasks

Task	Input	Output
Intention-to-Document (I2D)	Natural-language design intention (+ category, style, dimensions)	HTML/CSS document
Document Derendering (DD)	Document screenshot (+ dimensions)	HTML/CSS document
Element-to-Document (E2D)	Set of text and image elements (+ dimensions)	HTML/CSS document

DocHTML Dataset

DocHTML is built with an automated data synthesis pipeline: metadata collection → semantic annotation (InternVL3) → HTML/CSS code generation (Qwen3-Coder-480B) → image asset synthesis (FLUX.1-dev) → rendering (Playwright) → data cleaning.

Each sample provides:

• Metadata — document category, styles, dimensions, design intention, and a factual content description.
• HTML/CSS source code — a hierarchical, fully editable document.
• Synthesized image assets — generated for every <img> placeholder.
• Rendered screenshot — the document rasterized with Playwright.

Samples	265,206
Categories	111
Styles	32
Train / Val / Test split	8 : 1 : 1

The dataset is delivered as Parquet tables (with per-page-variation rows for the three task framings) plus sharded tarballs of rendered screenshots and per-instance image assets. The release also includes the model predictions for the paper's experiments (outputs of AnyDoc and the baselines on the benchmark splits) and the VLM-as-judge and derendering scores computed for those predictions. See DATASET.md for the full row schema, split definitions, media layout, and the predictions/scores tables.

⚠️ Image-asset restriction. The synthesized image assets were generated with FLUX.1-dev and carry an additional non-commercial restriction: they may not be used to train, fine-tune, or distill a model competitive with FLUX.1 [dev] / FLUX.1 Kontext [dev]. See FLUX_LICENSE_NOTES.md.

Method

General-purpose MLLMs produce low-quality documents. SFT on DocHTML yields strong document-generation capability but still exhibits content overflow, where elements extend beyond the specified document height. HARL addresses this with a height-aware reward computed from the ratio ρ = ĥ / h between the rendered height ĥ and the target height h — penalizing both overflow (ρ > 1) and underflow (ρ < 1 − γ) to prevent reward hacking.

Prompts

The model prompts used in the dataset construction pipeline and baselines are provided as plain text (with {...} placeholders) in prompts/:

• 01_intention_and_description_annotation.txt
• 02_html_css_code_generation.txt
• 03_baseline_task_instructions.txt
• 04_flux_i2d_baseline.txt

Note. This repository does not release training or inference code.

Accessing the DocHTML Dataset and Model Checkpoints

The DocHTML dataset and the AnyDoc model checkpoints (SFT + HARL) are available now under the Adobe Research License. To request access, email Christopher Tensmeyer at tensmeye@adobe.com with:

• Subject: Access to DocHTML Dataset
• Body: a statement acknowledging that you understand and agree to the terms of the Adobe Research License.

Once your request is received, you will be provided with download instructions for both the dataset and the model checkpoints.

By requesting and using these materials you agree to the terms of the Adobe Research License.

Citation

If you find AnyDoc or DocHTML useful in your research, please cite:

@InProceedings{Lin_2026_CVPR,
    author    = {Lin, Jiawei and Zhu, Wanrong and I Morariu, Vlad and Tensmeyer, Christopher},
    title     = {AnyDoc: Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {626-635}
}

License

The DocHTML dataset and AnyDoc model checkpoints are released under the Adobe Research License (noncommercial research use only).

The synthesized image assets are additionally subject to the FLUX.1 [dev] Non-Commercial License v1.1.1 — see FLUX_LICENSE_NOTES.md.