DesignLab: Iterative Slide Design

• DesignLab is a slide design system that iteratively refines presentation slides using a dual-model process for error detection and targeted correction.
• It employs a two-step workflow where a design reviewer flags layout inconsistencies and a design contributor corrects them, enhancing slide consistency.
• This approach supports interactive revisions, lowers design entry barriers, and achieves superior visual polish compared to traditional one-shot methods.

DesignLab is a slide design system grounded in an iterative detection–correction workflow that leverages two fine-tuned LLMs to incrementally refine presentation slides. By explicitly structuring the process around two roles—a design reviewer and a design contributor—the system emulates real-world collaborative and revisionary design practices. This approach enables DesignLab to improve upon typical single-step or rule-based slide generators, delivering slides with a visual polish and flexibility comparable to (or exceeding) leading commercial and research-based design-generation tools (Yun et al., 23 Jul 2025).

1. Iterative Detection–Correction Workflow

The central methodological innovation in DesignLab is the separation of slide refinement into an alternating cycle between a reviewer and a contributor, both instantiated as independently fine-tuned LLMs. The process is formalized as a sequence: $$D^{(i+1)} = \text{Contributor}\Bigl( \text{Reviewer}( D^{(i)} ) \Bigr)$$ where $D^{(0)}$ is a perturbed (rough) design draft encoded as structured JSON, and each subsequent revision is produced by reviewing and correcting the prior state. The reviewer inspects the draft and marks elements needing improvement with a TENTATIVE flag, while the contributor modifies or reinstates flagged elements to resolve all detected problems. Iterations continue until the reviewer produces no further TENTATIVE flags, or until a defined maximum number of steps is reached.

This iterative approach departs from the one-shot optimization typical of previous automated layout and slide-generation frameworks. It supports cumulative refinement: each loop can address new or remaining inconsistencies, such as minor misalignments, overwritten color attributes, or removed elements, allowing the slide to reach a level of quality previously unattainable in a single pass.

2. Fine-Tuned Model Roles and Training

DesignLab’s review–contribute decomposition relies on two distinct LLMs, each fine-tuned for its respective function:

• Design Reviewer: This model is instruction-tuned to take a JSON representation of a slide and detect elements that deviate from professional layout conventions. Detected elements are marked with a TENTATIVE label, signaling them as needing correction. Crucially, the review process is not based on a closed set of error classes but learns to flexibly identify a range of design issues, from spatial misalignments to font inconsistencies.
• Design Contributor: This model receives JSON with TENTATIVE annotations and is trained to update only those flagged elements. It adjusts properties (e.g., precise $(x, y)$ positions, RGB color values, font types) and may also reintroduce omitted elements. The contributor therefore restores or enhances the draft specifically in response to detected issues, minimizing unnecessary changes.

Training uses paired data: starting from a high-quality slide, controlled perturbations are introduced to generate rough drafts (see next section), and the models are trained on pairs of (perturbed, ground-truth) designs with supervision indicating the required detection or correction action.

3. Controlled Draft Perturbations and Simulation of Design Evolution

In absence of large-scale collections of real rough–final slide pairs, DesignLab simulates the design evolution process using controlled perturbations applied to well-formed slide JSONs. Perturbations include:

• Random removal or duplication of shapes.
• Displacement (offset) of element positions.
• Rotation and skew transformations.
• Color desaturation or replacement with a fixed set (e.g., Arial, Roboto, Calibri for fonts).
• Text attribute variation (size, font).

By carefully varying perturbation types and magnitudes, the system produces drafts that mirror the imperfections typical of early-stage designs by non-experts. These synthetic examples are then used to train the reviewer to reliably identify faults and the contributor to correct them, resulting in a system robust to a variety of real-world mistakes.

4. Performance Evaluation and Comparative Results

DesignLab’s effectiveness is assessed both qualitatively and quantitatively. Metrics include:

• Reviewer Detection Quality: Assessed by precision and recall in finding perturbations. For example, in experiments, reviewer precision for shape corrections reaches up to 0.87, with responsiveness near 1.0.
• Contributor Responsiveness: Measured as the fraction of flagged elements that are properly updated in the following iteration.
• Aesthetic Quality: Evaluated via user studies and through automated pairwise GPT-4o evaluations. Slides refined through the DesignLab process received consistently higher ratings (1–10 scale) and were repeatedly preferred in blinded comparisons to both prior research baselines and PowerPoint Designer.
• Branching and Diversity: The iterative looping enables not only progressive improvement but also the exploration of multiple design "branches," allowing users to select among alternative refinements.

The iterative methodology yields final slide ensembles with a higher degree of polish and fewer lingering imperfections than those generated by one-step approaches or commercial AI-assisted solutions.

5. Practical Implications and User Impact

DesignLab has clear benefits for non-expert and everyday users, directly addressing several challenges present in real-world slide creation:

• Lowering Entry Barriers: By automating pinpoint detection and correction, the system lessens the need for design expertise; non-specialists can iteratively enhance a slide without mastering every aspect of layout or visual hierarchy.
• Interactive Refinement: Users can interactively select specific objects to prioritize for refinement, compare branches, and incorporate personalized feedback during iterations.
• Efficiency and Accessibility: The system is hardware-efficient, operating with less than 8GB VRAM and sub-minute per-cycle latency (as low as 6s per iteration with optimization), rendering it suitable for general use even on consumer GPUs.
• Stability to Draft Quality: The design reviewer–contributor duo, trained on diverse perturbation patterns, imparts robustness against a wide variety of draft deficiencies and uncertainties.

6. Technical Summary and LaTeX Formalization

The iterative cycle central to DesignLab is mathematically represented as: $D^{(0)} = \text{Initial Perturbed Design (JSON)}$

$$D^{(i+1)} = \text{Contributor}\Bigl(\text{Reviewer}\bigl(D^{(i)}\bigr)\Bigr)$$

$$\text{Stop when } \text{Reviewer}\bigl(D^{(i+1)}\bigr) = \emptyset$$

The reviewer’s detection accuracy is quantified using: $$\text{Precision} = \frac{\text{True Positives}}{\text{True Positives + False Positives}}$$

$$\text{Recall} = \frac{\text{True Positives}}{\text{True Positives + False Negatives}}$$

where true positives are elements correctly labeled as TENTATIVE, and false positives/negatives are mislabelings against ground truth.

7. Broader Context and Outlook

By explicitly integrating error detection and targeted correction as iterative, independently learned model roles, DesignLab advances the state of automated slide design. Relative to prior art, which focuses on end-to-end layout optimization or static recommendation, this framework enables progressive improvement, supports personalized user-in-the-loop workflows, and yields slides that meet higher standards of visual polish and professional consistency.

DesignLab's methodology and results demonstrate the advantages of role-specialized and stepwise refinement in complex design environments, with broader implications for future systems targeting automated, interactive document and graphic design generation (Yun et al., 23 Jul 2025).