InfoVis: Unified Slide Image Generation and Editing

AI summary

Overview

Research area: Computer Vision, specifically conditional image generation and editing applied to presentation slides (structured, text-rich visual content).

Technical level: Advanced. The paper assumes familiarity with diffusion-based image generators, continual pretraining, supervised fine-tuning, multi-task training, and weight-space model merging, though the task framing itself is easy to follow.

Scope: The paper formulates four slide-centric generation and editing tasks, builds a large slide-native data engine and a diagnostic benchmark, and trains a unified 9B model (InfoVis) that generates and edits single-slide images.

What This Paper Is About

Generating a good presentation slide requires one system to render dense text accurately, follow design instructions, preserve supplied charts and images, arrange everything into a coherent layout, and apply local edits without disturbing unrelated content. Existing image generators handle only subsets of these abilities, and existing benchmarks mostly evaluate whole presentations or programmatic deck generation rather than individual slide images. The paper's goal is to close that gap by building slide-specific training data, a task-specific benchmark, and a single unified model that covers all four slide tasks.

Key Contributions

1. Task formulation and data engine. The authors define four tasks under a shared text-and-image interface — text-to-slide generation, template-conditioned generation, multi-image grounded generation, and instruction-based slide editing — and build a slide-native data engine producing over 13M continual-pretraining records and 1.9M supervised fine-tuning records.

2. VisSlideBench. A benchmark of 1,527 manually screened instances (943 in VisSlideBench-Gen and 584 in VisSlideBench-Edit) covering all four tasks with task-specific, multidimensional evaluation. The paper reports that its automatic dimension-wise rankings agree with blinded dimension-wise human rankings.

3. InfoVis, a unified 9B model. Trained from FLUX.2-klein-base-9B, InfoVis leads every evaluated open-source model on all eleven VisSlideBench metrics and scores 7.94 overall against 7.99 for Nano-Banana-2 and 8.36 for GPT-Image-2.

4. External validation and a workflow. InfoVis is the best open-source model in the BizGenEval Slides comparison, and a document-to-slide workflow built on it ranks second among ten systems in Aesthetics on SlidesGen-Bench.

Main Findings

• Slide-specific post-training, not scale, drives the gains. Starting from FLUX.2-klein-base-9B, slide-specific training improves the overall score by 1.64 points. The largest gains are in Content Faithfulness (+1.70 to +2.14), followed by Aesthetic Quality (+1.60 to +1.83) and Instruction Following (+0.85 to +1.02). Editing Semantic Consistency rises by 3.15 points. InfoVis, trained from a 9B base, outperforms the 32B FLUX.2-dev.

• Unification costs little. InfoVis uses one model for all four tasks and leads every open-source baseline, specialist or unified, on all eleven metrics. Against the authors' own task-specific checkpoints, the fused model loses only 0.06 on editing SC (9.13 versus 9.19) and improves the remaining metrics. By contrast, other unified open-source models such as BAGEL, OmniGen2, and HiDream-O1-Image score 4.1–4.9 overall and trail the editing specialist Qwen-Image-Edit on editing.

• The remaining gap to proprietary models is ordered by subjectivity. Against GPT-Image-2, InfoVis is closest on Content Faithfulness (gaps of 0.05–0.22), then Instruction Following (0.42–0.56), then Aesthetic Quality (0.55–0.69). InfoVis matches or exceeds Nano-Banana-2 on Content Faithfulness across all three generation tasks and achieves comparable Instruction Following, but trails on Aesthetic Quality. Editing is strong on localized operations such as text, color, and typography edits, while theme switching and other global transformations remain harder.

• Stage ablation shows non-uniform transfer. Text-to-slide continual pretraining alone raises T2I Content Faithfulness from 6.29 to 8.00 but collapses editing Semantic Consistency from 5.98 to 1.60. Four-task mixed continual pretraining restores all four groups and lifts the overall score to 7.38. Shared SFT reaches 7.73, and each individual task-specific SFT lowers the overall score below that shared checkpoint (template SFT drops editing SC to 6.49). Weighted averaging keeps nearly the peak of each specialization, giving the final 7.94.

• Human validation of the evaluator. Two groups of two domain experts each produced consensus rankings of five anonymized model outputs per example across all four tasks, yielding 2,153 valid pairwise comparisons. The GPT-4.1-based automatic evaluator reaches a mean Kendall's τb of 0.640, a mean Spearman's ρ of 0.700, and pairwise agreement of 88.71%.

• Test-time scaling helps most where the verifier is reliable. Best-of-8 selection with a Qwen3.5-27B slide-quality verifier on a random subset of 50 T2I cases raises the overall score from 7.96 to 8.11, with the largest gain in Text Rendering (8.57 to 8.80). Instruction Following and Aesthetic Quality improve more modestly and not monotonically.

• Structured prompts separate the models. On the text-to-slide split (115 short requests, 109 detailed descriptions, 93 StyleJSON specifications), InfoVis scores 8.00, 8.08, and 7.28, trailing GPT-Image-2 by 0.40 to 0.50 in every category. From detailed descriptions to StyleJSON, GPT-Image-2, Nano-Banana-2, and InfoVis each lose under 0.9 points, whereas open-source baselines lose 1.5 to 4.2 points.

• External OOD results. On BizGenEval Slides, InfoVis reaches 46.0/79.8 on the hard/easy splits — best among open-source models in that comparison, but below GPT-Image-2.0 (84.0/95.0) and Nano-Banana-2.0 (73.8/95.8); the paper notes Nano-Banana-Pro and Seedream-5.0, omitted from the table, also outperform InfoVis. On the SlidesGen-Bench Aesthetics track, InfoVis plus the document-to-slide workflow scores 27.09, close to the best reported result of 27.28 (Skywork-Banana), ranking second among ten systems. Only the Aesthetics track was run; content and editability were not.

Methodology in Plain English

The authors treat slide creation as one unified problem: produce an image from a text instruction plus optional input images. Depending on the task, those input images are absent (text-to-slide), a reference slide that sets the visual style (template-conditioned), one or more charts or figures that must appear in the output (multi-image grounded), or the slide to be modified (editing).

To get training data, they assemble a corpus of over 72K decks and 2.2M slide pages from author-released conference talks and presentation-sharing websites spanning 39 topical categories. They filter low-quality decks with Qwen3.5-27B, parse pages with PaddleOCR-VL into reading-order text plus localized figures and tables, then have a VLM write three kinds of instructions for each slide: a short caption, a detailed description, and a structured style_json layout specification. The same target slide therefore yields text-to-slide, template-conditioned, and multi-image records.

Editing data is built differently: the authors programmatically modify PPTX files using seven operations (text changes, block addition and removal, background, color, font, and typography changes), render before and after images, and ask an LLM to describe the change that actually occurred. Two further operations — theme switching and image replacement — are generated by Nano-Banana-2 and reserved for the fine-tuning stage.

For supervised fine-tuning, Nano-Banana-2 renders roughly 300K target slides from LLM-generated concepts, real SlideShare content, and paper-derived specifications, which are then annotated. The counts reported are about 890K T2I records, about 300K template records, about 580K multi-image records, and about 140K editing records.

Training proceeds in three stages. First, continual pretraining on large-scale text-to-slide data to adapt the base generator to slides. Second, a four-task mixed continual pretraining phase. Third, one epoch of text-to-slide SFT, then four task-specific fine-tunes from the resulting shared checkpoint, which are linearly fused by weight-space merging. The base architecture is kept; the maximum prompt length is raised to 4,096 tokens.

Evaluation uses OCR for text rendering and GPT-4.1 as a judge for visual content, design-instruction compliance, and presentation quality. Content Faithfulness uses a normalized edit distance adapted to the required text. Instruction Following is scored on Background Compliance, Color Palette Adherence, Typography Compliance, Layout and Position Accuracy, Mood and Atmosphere Match, and Block-Level Style Fidelity. Aesthetic Quality is judged from the output image alone on Clarity, Layout, Style Coherence, and Presentation Professionalism. For editing, the evaluator scores Instruction Following and Edit Consistency, with Semantic Consistency defined as their minimum and the overall editing score as the square root of Semantic Consistency times Presentation Quality.

Why This Matters

Impact on research. The paper argues that the bottleneck for slide generation is not model scale but the absence of slide-native supervision and diagnostic evaluation. It supplies both a data engine and a benchmark whose automatic rankings are validated against human judgments, giving the field a reproducible way to isolate individual slide capabilities such as text fidelity, template consistency, visual grounding, and edit preservation.

Real-world applications:

• Automatically converting a research paper or business report into a well-designed slide deck.
• Generating on-brand slides from a template reference so that recurring decks keep consistent background, palette, and typography.
• Producing slides that must incorporate specific charts, tables, or figures supplied by a user without altering them.
• Making targeted edits to an existing slide — changing text, recoloring, swapping fonts, replacing an image — while leaving everything else intact.

Industry relevance. Presentation creation is a common knowledge-work task, and the paper's external evaluations target commercial settings: the BizGenEval Slides domain and the agentic SlidesGen-Bench comparison against systems such as NotebookLM, Gamma, and Kimi. The finding that a 9B slide-tuned model can beat a 32B general-purpose generator, and that one unified model can match task specialists, is directly relevant to teams deciding how to allocate compute and whether to deploy separate models per task.

Future Directions

• Close the aesthetic gap. The remaining distance to proprietary models is largest in aesthetic quality and instruction following, and the paper attributes it to visual organization and presentation polish rather than content fidelity.
• Improve global edits. Localized operations such as text, color, and typography edits work well, but theme switching and other global transformations remain more difficult.
• Target knowledge- and reasoning-intensive cases. The gap on the BizGenEval hard split is attributed to cases that slide-specific training does not address.
• Address optimization desiderata that the paper itself identifies. Each individual task-specific SFT lowers overall performance, partly from cross-task interference and partly from distribution shift toward synthetic SFT pools; the paper notes multi-image SFT lowers its own task from 8.50 to 8.12 CF, mostly on cases built from real decks. Better fusion or data-balancing strategies are left open.
• Run the remaining SlidesGen-Bench tracks. Only the Aesthetics track was evaluated; content and editability were not run, leaving those comparisons unresolved.

Target Audience

Researchers and engineers working on text-rich or design-oriented image generation, unified generation-and-editing models, and multimodal training data construction. It is also useful for practitioners building document-to-slide or presentation-automation products, and for benchmark designers interested in how task-specific, dimension-wise evaluation is constructed and validated against human rankings. Readers without a background in diffusion model training will find the task framing and benchmark sections accessible, but the training and ablation sections assume familiarity with fine-tuning and model merging.

Authors’ abstract

Generating and editing presentation slides requires models to combine accurate text rendering, instruction following, visual grounding, aesthetic design, and precise editing within one canvas. Existing work covers only subsets of these capabilities and lacks large-scale slide-native supervision and diagnostic evaluation. We formulate four slide-centric tasks and build a scalable data engine with over 13M continual-pretraining records and 1.9M supervised fine-tuning records. We also introduce VisSlideBench, a benchmark of 1,527 manually screened instances that covers all four tasks with task-specific, multidimensional evaluation. Its automatic rankings agree with blinded, dimension-wise human rankings. Using these resources, we train InfoVis, a unified 9B slide generation and editing model. On VisSlideBench, InfoVis leads every evaluated open-source model on all eleven metrics and scores 7.94 overall against 7.99 for Nano-Banana-2. Starting from FLUX.2-klein-base-9B, slide-specific post-training adds 1.6 points overall, with the largest generation gains in content faithfulness; the remaining gap to proprietary models is smallest in content faithfulness and largest in aesthetic quality. Built on InfoVis, our document-to-slide workflow achieves the second-highest Aesthetics score on the external SlidesGen-Bench. We will release the model, code, data, and benchmark.