Design First, Code Later

Aesthetically Pleasing Template-Free Slides Generation

A hierarchical, design-first framework that decouples slide-level style design from page-level code implementation — no templates required.

---

Highlights

🏆	76.5% Top-1 human preference vs all open-source baselines — exceeding the second-best (9.5%) by 68.0 pp
🏆	52.0% win rate vs commercial systems (Kimi, Manus) — exceeding the second-best (19.5%) by 32.5 pp
📊	Highest VLM-judge average score 3.78 across Layout, Hierarchy, Color, Clarity, and Coherence
🎨	Human eval: Clarity & Structure 4.25–4.29 · Visual Design & Aesthetics 3.84–3.98

---

Comparison with Existing Methods

Method	Output	Full Deck	Template-Free
EvoPresent	HTML	✅	❌
AutoSlides	LaTeX/PDF	✅	❌
SlideGen / SlideCoder	PPTX	❌	❌
Kimi	PPTX/PDF	✅	✅
Manus	PPTX/PDF	✅	✅
DeepSlides (ours)	PPTX/Image	✅	✅

---

How it works

DeepSlides operates at two levels:

Slides-level Design — determines the global visual identity of the entire deck: tone, color palette, font colors, decorative shapes, and layout diversity guidelines. Pages are categorized into functional pages (cover, section dividers, end page) and content pages, each receiving tailored style instructions.

Page-level Generation — each slide is generated through four stages:

Content Expansion  →  Design  →  Implementation  →  Evaluation & Refinement
  (web search)       (3-layer)    (Python/PPTX)     (completeness · compliance · aesthetics)

The three design layers per slide:

• Background layer — textures, decorative elements, patterns
• Layout layer — block arrangement, spatial positions, structural boundaries
• Content layer — exact text snippets, images, visual elements

The full pipeline in graph.py:

Input (topic / image / style)
  │
  ├─[Image Analysis]          optional: infer research intent from a figure
  │
  ├─[Report Planning]         outline sections and search queries
  │
  ├─[Research & Writing]      parallel web search + writing per section
  │
  ├─[Slides-level Design]     global style, colors, shapes, tone
  │
  ├─[Cover / Chapter / End]   functional pages with matched styling
  │
  └─[Page-level Generation]   parallel per slide:
        expand → design (3 layers) → code → evaluate → refine
                                                            │
                                                            ▼
                                               presentation.pptx

---

Key Features

• Template-free — each slide's layout is designed from scratch to match its content, preventing visual fatigue across the deck
• Design–implementation decoupling — a designer module reasons in a semantic design space; a coder module translates specs into stable PPTX code, limiting error propagation
• Deep research integration — built on Open Deep Research, automatically retrieves web content, images, and academic sources per slide
• Three-dimension evaluation — each slide is scored on completeness, compliance, and aesthetics; low-scoring slides are iteratively refined before assembly
• Multi-model support — planner, writer, designer, and coder roles each use a configurable LLM (OpenAI, Azure, Anthropic Claude, or any OpenAI-compatible endpoint)
• Multi-search backend — Tavily · Perplexity · Exa · DuckDuckGo · arXiv · PubMed · Google Search · LinkUp
• Image input — provide a figure or screenshot; a vision model infers the research direction automatically
• Parallel execution — slides within a section are generated concurrently via LangGraph's Send() API

---

Example Outputs

---

Key Files

File	Purpose
src/open_deep_research/graph.py	Main workflow graph (all nodes and edges)
src/open_deep_research/configuration.py	All configurable parameters and defaults
src/open_deep_research/prompts.py	Prompt templates for every stage
src/open_deep_research/state.py	TypedDict / Pydantic state schemas
src/open_deep_research/utils.py	Search backends, image captioning helpers
src/open_deep_research/run.py	Batch runner (CSV input)

---

Quick Start

1. Clone and install

git clone https://github.com/sxswz213/DeepSlides
cd DeepSlides

python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -e .
cd pptx_tools && pip install -e . && cd -

2. Configure environment variables

cp .env.example .env
# edit .env and fill in your keys

Minimum required:

OPENAI_API_KEY=sk-...
OPENAI_API_BASE=https://api.openai.com/v1   # or your proxy / Azure endpoint
TAVILY_API_KEY=tvly-...

All supported variables

Variable	Purpose
OPENAI_API_KEY	Primary LLM key
OPENAI_API_BASE	Custom base URL / proxy
TAVILY_API_KEY	Tavily search
EXA_API_KEY	Exa search
GOOGLE_API_KEY / GOOGLE_CX	Google Custom Search
AZURE_OPENAI_API_VERSION	Azure OpenAI version
AZURE_OPENAI_DEPLOYMENT	Azure deployment name
CODER_API_BASE	Separate endpoint for coder model
DESIGNER_API_BASE	Separate endpoint for designer model
LANGCHAIN_API_KEY	LangSmith tracing (optional)

3. Run

Option A — Interactive UI (LangGraph Studio)

langgraph dev --no-reload
# the Studio URL will be printed in the terminal output

Fill in Topic, Presentation Minutes, Style in the input form and click Submit. The left panel shows the live workflow graph; the right panel streams execution logs per node.

Option B — Batch runner (CSV)

python src/open_deep_research/run.py

Edit csv_path, start, and max_rows in run.py. The CSV must have a Topic column; optional columns: image_path, style, presentation_minutes.

---

Configuration Reference

All parameters are in Configuration (configuration.py):

Parameter	Default	Description
planner_model	openai/gpt-4o-mini-2024-07-18	Report outline generation
writer_model	openai/gpt-4o-mini-2024-07-18	Section content writing
coder_model	anthropic/claude-haiku-4.5	PPTX code generation
designer_model	anthropic/claude-haiku-4.5	Slide layout design
search_api	tavily	Search backend
max_search_depth	5	Max search iterations per section
number_of_queries	2	Search queries per section
number_of_queries_for_ppt	1	Extra queries during slide enrichment

Override via configurable dict:

from langchain_core.runnables import RunnableConfig

config = RunnableConfig(configurable={
    "planner_model": "openai/gpt-4o",
    "search_api": "exa",
    "max_search_depth": 3,
})
result = asyncio.run(graph.ainvoke(state, config=config))

---

Dependencies

langgraph · langchain-openai · langchain-anthropic    # LLM / graph engine
python-pptx (via pptx_tools)                          # PPTX rendering
tavily-python · exa-py · duckduckgo-search            # search backends
google-cloud-vision                                   # optional image captioning
langsmith                                             # optional tracing

Optional system dependency for PNG export:

brew install libreoffice      # macOS
sudo apt install libreoffice  # Ubuntu

---

License

MIT © 2026 DeepSlides Authors

---

Citation

If you find this work useful, please cite our paper (PDF):

@inproceedings{cui2026design,
  title     = {Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation},
  author    = {Cui, Zhiyao and Wang, Chenxu and Hu, Shuyue and Zhang, Yiqun and
               Shao, Wenqi and Zhang, Qiaosheng and Wang, Zhen},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}