Research map / Image generation / Autoregressive and latent generation
AutoPresent: Designing Structured Visuals from Scratch
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SlidesBench, a benchmark for natural-language-to-slide generation, and AutoPresent, an 8B LLaMA-3.1 model LoRA-tuned on 7k instruction-program pairs that generates slides via python-pptx code using a SlidesLib helper library.
Problem
Automated generation of structured visuals such as presentation slides from natural language instructions, requiring content creation and visual planning, which lacked benchmarks and evaluation metrics.
Approach
Frame slide generation as NL-to-program synthesis executed via python-pptx, with a high-level tool library (SlidesLib) for text, layout, and image search/generation, and train a small open model on program-annotated slide data; add self-refinement using rendered slide snapshots.
Key results
- AutoPresent (8B) achieves overall 79.0 execution rate and 55.2 overall score vs GPT-4o 58.0 on detailed instructions with images setting with SlidesLib (reported 55.0 vs 58.0 in Table 3)
- LLaMA 3.1 8B and LLaVA barely produce executable slides without fine-tuning (2.1% and 11.3% execution in detailed-with-images setting without library)
- One round of self-refinement with GPT-4o raises overall scores in all three settings (e.g., 58.0 to 59.5 in detailed+images)
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoPresent with SlidesLib instead of AutoPresent without SlidesLib (ablation) | AutoPresent · Detailed instructions with images | overall score: 45.2 -> 55.0 | Helped |
| Add the expert-designed SlidesLib library | GPT-4o · Detailed instructions only | overall score: 28.7 -> 56.3 | Helped |
| Add the expert-designed SlidesLib library | GPT-4o · High-level instructions | overall score: 39.7 -> 58.5 | Helped |
| Add one round of iterative refinement | GPT-4o with SlidesLib · Detailed instructions with images | overall score: 58.0 -> 59.5 | Helped |
| Add one round of iterative refinement | GPT-4o with SlidesLib · Detailed instructions only | overall score: 56.3 -> 59.5 | Helped |
| Add one round of iterative refinement | GPT-4o with SlidesLib · High-level instructions | overall score: 58.5 -> 59.8 | Helped |
| Continue iterative refinement beyond the first round (ablation) | GPT-4o with SlidesLib · Detailed instructions with images | overall score: 59.5 -> 59.3 -> 60.1 | Mixed |
| Continue iterative refinement beyond the first round (ablation) | GPT-4o with SlidesLib · Detailed instructions only | overall score: 59.5 -> 60.1 -> 59.4 | Mixed |
| Continue iterative refinement beyond the first round (ablation) | GPT-4o with SlidesLib · High-level instructions | overall score: 59.8 -> 61.3 -> 61.4 | Helped |
| Use code generation instead of end-to-end image generation | Slide generation methods · Detailed instructions only | overall score: DALLE 3: 50.2; GPT-4o with SlidesLib: 56.3 | Mixed |
Surprising findings
- Small open models like LLaMA 8B and LLaVA 7B can barely produce executable slides out of the box, despite strong text generation ability.
- End-to-end image generation (DALL-E 3, Stable Diffusion) produces creative visuals but often invalid or meaningless text and poor slide structure.
Limitations
- Focuses on single-slide generation rather than full slide decks
- Generates full slide code in a single pass without iterative interactive workflows
- Reference-free metrics rely on GPT-4o judgments, with human agreement measured only via ICC
Experimental setup
Datasets: SlidesBench training set (7k examples from 300 slide decks)
Benchmarks: SlidesBench
Metrics: element match content similarity color similarity position similarity reference-free text/image/layout/color scores execution success rate
Baselines: GPT-4o LLaMA 3.1 8B LLaVA v1.5 7B CodeLlama 7B DeepSeekCoder 7B Stable Diffusion 2 DALL-E 3
Builds on: visual programming Design2Code metrics LLaMA-3.1-8B-Instruct LoRA
Closest papers by meaning
- Generating Illustrated Instructions (CVPR 2024 · similarity 0.87)
- IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.86)
- OmniGen: Unified Image Generation (CVPR 2025 · similarity 0.86)
- From Words to Structured Visuals: A Benchmark and Framework for Text-to-Diagram Generation and Editing (CVPR 2025 · similarity 0.86)
- ShowHowTo: Generating Scene-Conditioned Step-by-Step Visual Instructions (CVPR 2025 · similarity 0.86)
- Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.