The Problem with Text-Only Generation

Traditional document-to-slide generation models often struggle with visual coherence and layout accuracy because they rely on text-based generation prompts. These models lack a feedback loop that accounts for how the final slide will actually render, leading to issues like overflow, poor visual hierarchy, and misalignment between content and design constraints.

The ReDeck Approach: Step-Level Grounding

ReDeck introduces a novel refinement framework that bridges the gap between abstract content generation and concrete visual rendering. Instead of generating a slide in a single pass, ReDeck employs a step-level, render-grounded refinement process:

• Iterative Feedback: The system generates an initial slide layout and then renders it. It then uses the visual output as grounding data to identify specific layout errors.
• Render-Grounded Correction: By analyzing the rendered output, the model can pinpoint issues such as text truncation, whitespace imbalances, or font sizing problems that are invisible to pure text-based LLMs.
• Step-Level Refinement: The model performs targeted adjustments at each step of the generation process, ensuring that the final slide adheres to both the source document's semantic structure and the visual constraints of the slide format.

This approach significantly improves the quality of automated slide decks by treating the rendering engine as a critical component of the reasoning loop, rather than just an end-stage output tool.