InfoVis: Unified Slide Image Generation and Editing

1 INTRODUCTION

• Existing benchmarks and general-purpose training do not systematically isolate single-slide capabilities or provide slide-native supervision for content, layout, and design.Slide generation also requires accurate text, coherent layouts, faithful use of visual assets, and edits that preserve unrelated content.Source: §1 · p. 1
• The paper formulates four tasks under a shared text-and-image interface: text-to-slide, template-conditioned, multi-image grounded generation, and instruction-based editing.Visual inputs serve as style references, content assets, or source slides to edit.Source: §1 · p. 2
• The slide-native data engine contains over 13M continual-pretraining records and 1.9M supervised fine-tuning records.It combines real presentations and synthetic supervision across the four tasks.Source: §1 · pp. 2–3
• VisSlideBench contains 1,527 manually screened instances and evaluates all four tasks with task-specific multidimensional measures whose dimension-wise rankings agree with blinded human judgments.Generation is evaluated for Content Faithfulness, Instruction Following, and Aesthetic Quality; editing for Semantic Consistency and Presentation Quality.Source: §1 · pp. 2–3
• 7.94 overall, InfoVis leads every evaluated open-source model on all eleven VisSlideBench metrics, versus 7.99 for Nano-Banana-2.The remaining gap to proprietary models is smallest in content faithfulness and largest in aesthetic quality.Source: §1 · pp. 2–3

Structured Rich-Text and Design Image Generation

• Prior design-image methods address complete artifacts, while general-purpose and unified generators lack slide-native supervision for template conditioning, multi-image grounding, and instruction-based editing.The paper positions InfoVis as a slide-centric generator with shared content, layout, and design supervision.Source: §2 · p. 3

Benchmarks for Slide Generation

• Existing benchmarks focus on programmatic slides, full presentations, broader visual content, or agentic workflows, while VisSlideBench targets single-slide generation and editing with task-specific measures.Its measures cover text fidelity, template consistency, visual grounding, and edit preservation.Source: §2 · p. 3

3 TASK FORMULATION AND SLIDE DATA ENGINE

• The data engine recovers slide content and visual elements as task-specific conditions while retaining the complete slide as the target; editing data comes from controlled changes to editable slides.The four tasks share content, layout, and design supervision.Source: §3 · p. 3, Figure 2 · p. 4

3.1 TASK FORMULATION

• Slide creation is formulated as y = Gθ(x, I): text-to-slide uses no image, template-conditioned generation uses a style reference, and grounded generation uses assets that must be preserved.For editing, I is the source slide and x is the requested change, with unrelated content kept stable.Source: §3.1 · p. 4
• Generation conditions include a short caption, a detailed content-and-design description, or structured style_json; editing uses a natural-language instruction.The same interfaces serve data construction and evaluation, aligning training tasks with VisSlideBench.Source: §3.1 · p. 4

3.2 CONTINUAL PRETRAINING DATA

• Real presentations are filtered and parsed into text in reading order and localized visual elements, which form a shared content representation for generation.The sources include author-released conference talks and presentations spanning 39 topical categories.Source: §3.2 · p. 4
• From parsed slides, generation records reuse the original rendered slide as target, with text-only, adjacent-slide template, or one-to-four-asset visual conditioning.Template pairs are retained when backgrounds, colors, typography, and decorative styles are consistent.Source: §3.2 · p. 5
• Continual-pretraining editing pairs are executable PPTX source–instruction–target triplets covering seven operations, filtered to remove defective slides, failed edits, and inconsistent instructions.The instructions describe changes realized by the executed PPTX actions.Source: §3.2 · p. 5

3.3 SUPERVISED FINE-TUNING DATA

• Generation SFT uses LLM concepts, SlideShare content, and paper-derived specifications, with Nano-Banana-2 rendering target slides and Qwen3.5-27B annotating records.Editing SFT selects high-quality executable triplets and adds model-generated theme-switching and image-replacement examples; the SFT resource contains about 1.9M records.Source: §3.3 · p. 5

4.1 BENCHMARK TASK SUITE

• VisSlideBench contains 1,527 manually screened instances across text-to-slide, template-conditioned, multi-image grounded generation, and editing; it jointly covers four creation interfaces and three prompt granularities.Its task-specific measures address content fidelity, instruction compliance, aesthetics, and visual grounding.Source: §4.1 · p. 5

4.2 BENCHMARK CONSTRUCTION

• Generation includes 493 real-presentation cases and 450 Nano-Banana-2-generated cases, while editing uses held-out cases from the editing-data pipeline and is therefore in-distribution.Manual screening removes poor-quality targets, ambiguous prompts, invalid pairs, imperceptible edits, and severe visual artifacts.Source: §4.2 · pp. 5–6

4.3 EVALUATION PROTOCOL

• Generation evaluation separates text rendering, instruction compliance, and output-only aesthetics; editing scores requested-change compliance and preservation of unrelated content, with Semantic Consistency defined as their minimum.Presentation Quality for editing uses the same four dimensions as generation Aesthetic Quality.Source: §4.3 · p. 6

4.4 HUMAN VALIDATION

• GPT-4.1’s dimension-wise rankings reach mean Kendall’s τb of 0.640, mean Spearman’s ρ of 0.700, and 88.71% pairwise agreement with expert consensus rankings.The validation covers 2,153 valid pairwise comparisons across all four tasks.Source: §4.4 · p. 6

5 INFOVIS TRAINING

• The model starts from FLUX.2-klein-base-9B, keeps its architecture, and raises the maximum prompt length to 4,096 tokens.Slide-native pretraining uses short captions, long descriptions, and structured design prompts.Source: §5 · pp. 6–7
• Table 1 compares T2I, Template, and Multi-image on content faithfulness, instruction following, and aesthetics, and Edit on semantic consistency and presentation quality.Overall averages the eleven metrics, with bold indicating the best result in each group.Source: Table 1 · p. 7
• InfoVis uses three training stages: slide-native continual pretraining, four-task mixed continual pretraining, and task-specific fine-tuning followed by model fusion.The first stage learns shared priors for text rendering, layout, visual hierarchy, and design; the second extends them to style-reference following, multi-image grounding, and editing.Source: §5 · p. 7
• The document-to-slide workflow extracts text and visual assets, plans slides and style prompts, renders candidates, and selects outputs with an external verifier.A selected slide serves as the template reference for subsequent pages.Source: §5 · p. 7

6.1 EXPERIMENTAL SETUP

• Comparisons use the shared VisSlideBench protocol, with each model evaluated on tasks supported by its official interface and overlong prompts truncated.The comparisons include proprietary and open-source models.Source: §6.1 · p. 7, §6.2 · p. 7

6.2 MAIN RESULTS ON VISSLIDEBENCH

• 7.94 overall follows slide-specific training’s 1.64-point gain from FLUX.2-klein-base-9B, with the largest generation gains in Content Faithfulness (+1.70 to +2.14).The other generation gains are +1.60 to +1.83 on Aesthetic Quality and +0.85 to +1.02 on Instruction Following.Source: §6.2 · pp. 7–8
• InfoVis leads every open-source specialist or unified baseline on all eleven metrics, losing just 0.06 on editing SC (9.13 versus 9.19) against task-specific checkpoints.The earlier shared checkpoint retains a slightly higher Template CF (8.25 versus 8.14).Source: §6.2 · p. 8
• Against GPT-Image-2, generation gaps are 0.05–0.22 on CF, 0.42–0.56 on IF, and 0.55–0.69 on AQ; global edits remain harder than localized operations.InfoVis matches or exceeds Nano-Banana-2 on Content Faithfulness across all three generation tasks, but trails on Aesthetic Quality.Source: §6.2 · p. 8

6.3 EXTERNAL BENCHMARK EVALUATION

• 46.0/79.8 on BizGenEval hard/easy makes InfoVis the best open-source model there; its document-to-slide workflow scores 27.09 Aesthetics versus 27.28 best on SlidesGen-Bench.Nano-Banana-Pro and Seedream-5.0 also outperform InfoVis on BizGenEval Slides; the SlidesGen-Bench evaluation reports only the Aesthetics track.Source: §6.3 · p. 8, Table 2 · p. 8, Table 3 · p. 8

6.5 MORE ANALYSIS

• Best-of-8 raises overall from 7.96 to 8.11, chiefly improving Text Rendering from 8.57 to 8.80.Instruction Following and Aesthetic Quality improve more modestly and not monotonically across dimensions.Source: Figure 3 · p. 9, §6.5 · p. 9
• InfoVis scores 8.00, 8.08, and 7.28 on short, detailed, and StyleJSON prompts, and remains ahead of clipped-window baselines when rescored on full inputs.It trails GPT-Image-2 by 0.40 to 0.50 on every prompt category.Source: Figure 3 · p. 9, §6.5 · pp. 9–10

6.4 ABLATION ON TRAINING STAGES

• T2I-only CPT raises T2I CF from 6.29 to 8.00 but drops editing SC from 5.98 to 1.60; weighted averaging yields the strongest unified model at 7.94 overall.Four-task mixed CPT restores all four groups, while task-specific SFT checkpoints each score below the shared checkpoint overall.Source: §6.4 · p. 9, Table 4 · p. 9

7 CONCLUSION

• InfoVis achieves the strongest overall open-source performance on VisSlideBench, while the document-to-slide workflow’s aesthetic quality is on par with the best reported systems.The framework, training resource, and benchmark are presented as the paper’s central contributions.Source: §7 · p. 10
• AI assistants supported writing, literature search, data annotation and filtering, SFT-target and benchmark-case synthesis, and automatic evaluation under author review.The authors state that all research ideas, claims, numbers, and final content were reviewed and confirmed by them.Source: AI USE STATEMENT · p. 10
• The authors will release InfoVis weights, code, training data, the data engine, and VisSlideBench for research use.Real-slide records will retain source information for tracing and removal on request.Source: REPRODUCIBILITY STATEMENT · p. 10, ETHICS STATEMENT · p. 10
• InfoVis can render inaccurate or misleading text, so the document-to-slide workflow should be used with its source document at hand; real-person content is not targeted.The synthetic data do not represent real individuals, and the human validation study collected no personal data from annotators.Source: ETHICS STATEMENT · p. 10

A.1 DATA SOURCES, FILTERING, AND PARSING

• The engine draws on over 72K decks and 2.2M pages from academic and 39 topical sources, filters decks for authenticity and quality, then parses retained pages into text and visual assets.Academic decks include substantial text, notation, equations, and data figures; topical decks add templates, color schemes, and graphic styles.Source: §A.1 · p. 14

A.2 GENERATION RECORD CONSTRUCTION

• Generation records cover text-to-slide annotations, adjacent-slide template conditioning with style checks, and multi-image asset grounding with placement instructions and crop filters.Text-to-slide and grounded-generation annotations include a short caption, detailed description, and structured style specification.Source: §A.2 · p. 14

A.3 EDITING PAIR CONSTRUCTION

• Seven executable PPTX operations and two model-generated edits form source–instruction–target pairs; actions are rendered, realized changes recaptioned, and pairs filtered.Theme switching and image replacement use model-generated targets; the remaining operations are constructed through PPTX editing.Source: §A.3 · p. 15

A.4 SUPERVISED FINE-TUNING RECORDS

• About 300K synthesized 16:9 targets feed about 890K T2I, 300K template, and 580K grounded SFT records, alongside about 140K editing records.The seven executable editing operations also contribute to CPT, while theme switching and image replacement are SFT-only.Source: §A.4 · p. 15, Table 5 · p. 16

A.5 DATA STATISTICS

• Table 6 counts instruction records by branch; multiple records can share a target slide or differ only in instruction or visual condition, and source deck/page counts precede filtering.Figure 4 breaks training records down by task and CPT/SFT pool, with generation source and domain splits and editing-operation splits.Source: Table 6 · p. 16, §A.5 · p. 16, Figure 4 · p. 17

B.1 COMPARISON WITH EXISTING BENCHMARKS

• Table 7 compares benchmark task coverage, evaluation criteria, and prompt formats; sizes use different units, and BizGenEval has 400 cases, including 80 slide cases.PPTBench combines generation with detection, understanding, and modification.Source: §B.1 · p. 16

B.2 BENCHMARK COMPOSITION

• Generation cases span three prompt forms and eight categories, while editing cases cover nine operations; benchmark decks, targets, and linked edit pairs are excluded from training.Figure 4 shows training distributions by task and CPT/SFT pool, with generation source/domain and editing-operation breakdowns.Source: §B.2 · pp. 16–17, Figure 4 · p. 17

B.3 EVALUATION DETAILS

• Content Faithfulness matches required strings against OCR-detected slide text using normalized edit distance scaled to 1–10, and averages text and image-preservation scores for multi-image generation.GPT-4.1 extracts required strings, while PaddleOCR detects and joins text lines; a separate GPT-4.1 judgment checks whether required input images are preserved and incorporated.Source: §B.3 · pp. 17–18

B.4 HUMAN VALIDATION PROTOCOL

• Human validation used blinded, randomized five-model comparisons with two-annotator consensus; 137 examples yielded 328 valid dimension rankings, analyzed with Kendall’s τb, Spearman’s ρ, and pairwise agreement.Annotators ranked task-relevant dimensions, with ties allowed; automatic-versus-human agreement is reported in Table 10.Source: §B.4 · pp. 19–20, Table 10 · p. 20

B.5 BENCHMARK EVALUATION PROMPTS

• The benchmark prompts cover instruction following for textual and structured styles, aesthetic quality, visual content faithfulness, editing success and overediting, and presentation quality.Per-dimension scores are averaged without using the instruction-following and aesthetic-quality templates’ overall_score; the image-preservation template’s overall_score is used directly.Source: §B.5 · pp. 20–21

B.5.1 INSTRUCTION FOLLOWING: TEXTUAL STYLE DESCRIPTION

• Textual-style instruction following is judged by strict, requirement-by-requirement comparison across background, palette, typography, layout, mood, and block-level styling, without rewarding visual attractiveness.The evaluator uses optional reference slides only for style context, not target-slide layout; uncertain or unverifiable requirements count as mismatches and scoring caps penalize deviations.Source: §B.5.1 · pp. 21–22

B.5.2 INSTRUCTION FOLLOWING: STRUCTURED STYLE JSON

• Structured-style instruction following applies the same conservative, requirement-by-requirement evaluation to Style JSON specifications, including global style, visual hierarchy, and block attributes.Optional reference slides provide style context only; requirements are marked PASS, PARTIAL, FAIL, or NOT_VERIFIABLE, with mismatches informing capped sub-dimension scores.Source: §B.5.2 · pp. 23–24
• Structured Style JSON compliance is scored by typography, layout and block-level fidelity; a sub-dimension earns 9–10 only when every checkable requirement passes with zero mismatches, while two or more mismatches cap it at 4.Source: §B.5.2 · pp. 24–25

B.5.3 AESTHETIC QUALITY

• Aesthetic Quality is judged by readability, layout, style coherence, and professional polish, with visible defects counted and conservative caps applied before averaging sub-dimension scores.The evaluator inspects the slide globally, identifies strengths and weaknesses, and does not let attractiveness compensate for structural or readability flaws.Source: §B.5.3 · pp. 26–28

B.5.4 CONTENT FAITHFULNESS: REQUIRED INPUT IMAGES

• Content Faithfulness requires checking every listed sub-image against the generated presentation for semantic content, visual identity, composition, and positional consistency.Each item is classified as EXACT_MATCH, PARTIAL_MATCH, FAIL, or NOT_FOUND; partial matches count as mismatches and not-found items as major mismatches before conservative scoring.Source: §B.5.4 · pp. 29–30
• Required input images are judged for presence, semantic correctness, and structural consistency, with strict score caps for missing items and major mismatches.The criteria reject superficial resemblance and prioritize semantic correctness over visual attractiveness; caps also apply when partial matches or worse exceed specified proportions.Source: §B.5.4 · pp. 30–32

B.5.5 EDIT EVALUATION: EDITING SUCCESS AND OVEREDITING

• The edit evaluator compares the source and edited slides against the instruction, scoring Editing Success and Overediting from 0 to 10.Editing Success measures whether the requested change is applied; Overediting measures whether only intended parts change while the rest is preserved.Source: §B.5.5 · p. 32

B.5.6 EDIT EVALUATION: PRESENTATION QUALITY

• Edited-slide presentation quality is scored from 0 to 10 on clarity, layout, style coherence, and presentation professionalism.The dimensions assess readability and hierarchy, composition and spacing, stylistic consistency, and whether the page looks polished and professional.Source: §B.5.6 · p. 33

C.1 EXPERIMENTAL SETUP DETAILS

• Training uses 13,500-step continual pretraining followed by 16,500-step four-task SFT, and the final model linearly interpolates four task checkpoints.The stages use global batch sizes of 384 and 640, respectively; inputs are capped at 4,096 tokens, and the Qwen3 text encoder and VAE remain frozen. Inference uses 50 denoising steps; verifier-guided best-of-K selection scores eight candidates without updating InfoVis.Source: §C.1 · pp. 33–34
• Paired bootstrap estimates InfoVis 0.06 below Nano-Banana-2 overall and 0.42 below GPT-Image-2 overall.The 95% confidence intervals are 0.01–0.10 and 0.38–0.46, respectively; there is no detectable difference on the nine generation metrics (CI −0.02 to 0.07). IF and AQ are unweighted means of per-dimension judge scores; the judge’s overall_score is not used for those metrics.Source: §C.1 · p. 34

C.2 DETAILED GENERATION AND EDITING RESULTS

• Detailed generation-task averages weight listed dimensions equally, unlike the eleven group-level metrics in the main comparison.Tables 12–14 report the three generation tasks, while Table 15 reports nine editing operations.Source: §C.2 · p. 34

• The benchmark includes 167 real and 150 synthetic T2I records, 147 and 150 template-conditioned records, and 179 and 150 multi-image records.Each task score averages CF, IF, and AQ within a source; the macro score weights the three task means equally.Source: §C.3 · pp. 34, 36
• InfoVis ranks first among open-source models on every task for both sources; its Real and Synthetic macro scores are 7.86 and 7.80, with a 0.060 gap versus the 0.068 mean gap.Source: Table 16 · p. 36, §C.3 · p. 36
• Real–Synthetic Spearman rank correlations are 1.000 for T2I, 0.983 for template-conditioned generation, 0.929 for multi-image generation, and 0.950 for the macro score.The table supports ranking robustness to source, not independence from it, because the partitions contain different slide content.Source: Table 16 · p. 36, §C.3 · p. 36

C.4 MULTI-IMAGE SFT BY BENCHMARK SOURCE

• On real cases, multi-image SFT lowers CF 8.64 to 7.99, IF 8.05 to 7.55, and AQ 6.68 to 5.94; oversized or redrawn assets can overlap surrounding text.Source: §C.4 · p. 36
• On synthetic cases, multi-image SFT lowers CF 8.34 to 8.29, IF 7.53 to 7.23, and AQ 6.97 to 6.67, while CF asset preservation rises from 8.03 to 8.29.Source: §C.4 · p. 36
• Template SFT lowers multi-image CF to 7.62 on real and 7.36 on synthetic cases; fusing checkpoints that retain real-slide priors yields the best score in every column.Source: Table 17 · p. 37, §C.4 · p. 37

C.5 PROMPT-LENGTH TRUNCATION

• A 1,024-token window clips every StyleJSON prompt (1,102–3,731 tokens), but no short or detailed prompts; truncation does not explain the StyleJSON score pattern.On prompts received in full, InfoVis remains ahead of three open-source baselines on all three metrics and ties Qwen-Image-2 on T2I; it remains the best open-source model on the 80 untruncated prompts.Source: Table 18 · p. 37, §C.5 · p. 37

• The workflow parses document text and visual assets, then plans an outline and per-slide content before determining how to arrange and style it.Source: D DOCUMENT-TO-SLIDE WORKFLOW · p. 37
• GPT-4.1 converts each slide content plan into a style_json prompt specifying text, hierarchy, image placement, typography, colors, background, and decoration for InfoVis to render.Source: D DOCUMENT-TO-SLIDE WORKFLOW · p. 38
• InfoVis generates K candidates per slide, and an external verifier scores them against the prompt to select the highest-scoring candidate.The selected first slide serves as a style reference for every later slide, maintaining visual consistency while each follows its own content plan.Source: D DOCUMENT-TO-SLIDE WORKFLOW · p. 38, Figure 7 · p. 38
• Optional PowerPoint export reconstructs recognized text as editable text boxes, while figures and tables remain cropped image assets.OCR and layout analysis recover text, positions, and basic styles; text regions are removed and inpainted to create a background image.Source: D DOCUMENT-TO-SLIDE WORKFLOW · p. 38

• VisSlideBench-Edit uses the same construction pipeline as editing training data, so it measures in-distribution editing quality; English-language slides are the only language covered.Source: E LIMITATIONS · p. 38
• Generation SFT targets and half of VisSlideBench-Gen are produced with a proprietary model, tying training and evaluation cases to its style.The paper reports real and synthetic generation results separately in Table 16.Source: E LIMITATIONS · p. 38
• Aesthetic Quality has the largest remaining gap to proprietary models, and knowledge- and reasoning-intensive slides fall outside the targets of slide-specific supervision.Source: E LIMITATIONS · p. 38
• The paper illustrates failures in visual-asset fidelity and in preserving content outside an editing target.Source: E LIMITATIONS · p. 38

F QUALITATIVE RESULTS

• InfoVis produces coherent text-based layouts, follows reference styles, integrates multiple visual assets, performs local and global edits, and generates visually consistent multi-slide presentations.The examples cover all four slide tasks and the document-to-slide workflow.Source: F QUALITATIVE RESULTS · p. 39

1 Visual asset fidelity

• Generated charts can retain their broad appearance while corrupting legend entries and axis labels and changing bar-group structure.The supplied chart is reconstructed rather than faithfully preserved.Source: G FAILURE CASES · p. 44, Figure 13 · p. 44

G FAILURE CASES

• An edit requesting changes only to two left-hand graphs instead produces four plots and removes an out-of-target table on the right.The model identifies what to change but does not reliably confine changes to the requested region.Source: G FAILURE CASES · p. 44, Figure 13 · p. 44