acceptodds
Under review as a conference paper at ICLR 2027

InfoVis: Unified Slide Image Generation and Editing

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 gains in content faithfulness; the remaining gap to proprietary models grows with subjectivity, from content faithfulness through instruction following to 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.

Primary area:
generative models
Keywords:
Generative Modeling, Unified Model, Image Editing, Text-to-Image Generation, Benchmark

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N8iNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper approaches PPT slide generation and editing as an image generation problem rather than a PowerPoint manipulation problem.
  • The paper formulates slide creation as a “unified conditional visual synthesis problem.”
Critical strengths
  • Lines 42–44: I agree with the authors’ observation that the image distributions of general image generators and presentation slides are different.
  • The paper explores slide generation and editing from an image-generation perspective, rather than treating them solely as PowerPoint manipulation tasks. This is an interesting direction.
Critical weaknesses
  • Lines 46, 135–146: The discussion of slide benchmarks should cite relevant prior work, including Talk to Your Slides: High-Efficiency Slide Editing via Language-Driven Structured Data Manipulation and EditPPT: Faithful Long-Deck Slide Editing via Structured....

  • Line 49: The paper describes slide creation as a “unified conditional visual synthesis problem,” but it is unclear what exactly is being unified.

  • I am not sure whether editing generated slide images should be referred to as “slide editing.” I think “image editing” may be a more appropriate term. “Slide editing” sounds more like editing individual objects within a PowerPoint slide.

  • Line 88: The paper itself refers to some operations as “deterministic PPTX transformations.” Given this, I wonder how redundant the slide-native data engine is in practice. It would be helpful to provide statistics showing how often or to what extent this component is actually needed.

  • In PPT slide editing, iterative editing is inevitably needed. For example, with Nano-Banana-2, repeated iterations in PowerPoint slide generation can cause text to become corrupted or image quality to gradually degrade. I wonder whether the proposed model exhibits similar behavior.

  • I am not suggesting that the authors necessarily need to eliminate this phenomenon, but I believe such behavior, if present, should be reported as a limitation of the paper.

  • Section 6.5: The title “More Analysis” sounds somewhat awkward. It may be better to separate the two analyses into Sections 6.5 and 6.6, or use a title that captures what the two analyses have in common.

  • The most critical weakness is the evaluation setup. Evaluating the authors’ main system primarily on a benchmark created by the authors themselves weakens the credibility of the evaluation.

  • The proposed method should also be sufficiently evaluated on the other existing benchmarks cited in the paper.

  • Figure 5: Looking closely at the slide example in the Generation Evaluation box, the rightmost 3×5 input image becomes a generated slide with labels (a), (b), ..., (e) for the columns.

  • The most noticeable weakness of the generated slide in Figure 5, in my view, is that the Figure 5 caption is center-aligned rather than justified. However, the current benchmark structure does not appear to evaluate this type of formatting issue.

  • Evaluation criteria for slides may vary considerably depending on the slide and the evaluator. This raises concerns about whether rubric-based evaluation alone is sufficient.

  • OCR results are unlikely to contain detailed information about the precise locations of text. Therefore, for slides where text placement is important, it is unclear whether OCR-based evaluation is truly effective.

  • There also appears to be substantial overlap between the OCR-based evaluation and the GPT-4.1-based evaluation. GPT-4.1 may already cover many of the aspects being assessed through OCR, and potentially additional aspects as well.

Questions for authors
  • Line 49: The paper formulates slide creation as a “unified conditional visual synthesis problem.” What exactly is being unified in this formulation?

  • Do the authors consider editing a generated slide image to be “slide editing” rather than “image editing”? Since “slide editing” usually suggests manipulating individual objects within a PPT/PPTX file, I would appreciate clarification on the terminology.

  • Line 88: Given that the paper itself refers to certain operations as “deterministic PPTX transformations,” how redundant is the slide-native data engine in practice? Do the authors have statistics quantifying how often this component is required?

  • Does the proposed model suffer from degradation under repeated iterative editing? For example, repeated edits with Nano-Banana-2 in PowerPoint slide generation can lead to corrupted text or progressively degraded image quality. If similar behavior occurs in the proposed system, could the authors report it as a limitation?

  • More fundamentally, is treating PPT slide generation and editing as an image-generation problem really the right direction?

  • Systems such as OpenAI’s GPT and Claude can directly manipulate PowerPoint applications and support detailed editing. Compared with this direction, what are the advantages of solving slide generation and editing through image generation, and why is this formulation preferable?

  • Slides are fundamentally designed to be presented and edited by people in tools such as Google Slides or PowerPoint. They may also include animations in which individual objects appear sequentially. Since the proposed image-based formulation cannot represent such characteristics, how do the authors view this limitation?

  • Figure 5: Were the (a), (b), ..., (e) labels in the generated slide explicitly specified in the instruction? If so, does the evaluation framework verify whether such details are correctly generated?

  • If alignment and other detailed layout issues, such as the center-aligned caption in Figure 5, are not captured by the current benchmark, how can the benchmark adequately evaluate slide quality when the criteria may vary across slides and evaluators?

  • For slides in which the exact position of text is important, how effective is OCR-based evaluation, given that OCR results may not capture detailed spatial information?

  • Would it be more appropriate to evaluate these aspects directly using a VLM such as GPT-4.1? Since GPT-4.1 is already used in the evaluation, what is the motivation for separately using OCR, and what does OCR measure that is not already captured by GPT-4.1?

I reviewed the paper without using AI, so there may be points I have missed. I will raise my score if the authors provide sufficient and convincing explanations. Thank you.

8 comments

We thank the reviewer for the careful, line-by-line review and for the constructive comments on terminology, iterative editing, external evaluation, the evaluation protocol, and the image-based formulation. We number the weaknesses (W1–W14) and questions (Q1–Q11) in the order they appear in the review and address each point below; all changes in the revised paper are marked in blue, and the new experiments are in the new Appendix H; table numbers refer to the revised paper.

W1: Citations

W1. Lines 46, 135–146: The discussion of slide benchmarks should cite relevant prior work, including Talk to Your Slides: High-Efficiency Slide Editing via Language-Driven Structured Data Manipulation and EditPPT: Faithful Long-Deck Slide Editing via Structured....

We now cite and discuss Talk to Your Slides (Jung et al., 2026) and EditPPT (Kim et al., 2026) in Section 2 and add them to Table 8.

W2, Q1: What is unified

W2. Line 49: The paper describes slide creation as a "unified conditional visual synthesis problem," but it is unclear what exactly is being unified.

Q1. Line 49: The paper formulates slide creation as a "unified conditional visual synthesis problem." What exactly is being unified in this formulation?

"Unified" refers to two things. First, the task formulation: all four tasks share one interface, y = Gθ(x, I), and differ only in the role of the optional images I, which are absent in text-to-slide generation and are the template slide image, the embedded asset images, and the slide image to be edited in the other three tasks. Second, the model: a single set of model weights performs well on all four generation and editing tasks. We have revised the paragraph at line 49 to state this explicitly.

W3, Q2: Terminology

W3. I am not sure whether editing generated slide images should be referred to as "slide editing." I think "image editing" may be a more appropriate term. "Slide editing" sounds more like editing individual objects within a PowerPoint slide.

Q2. Do the authors consider editing a generated slide image to be "slide editing" rather than "image editing"? Since "slide editing" usually suggests manipulating individual objects within a PPT/PPTX file, I would appreciate clarification on the terminology.

We agree. Our model edits rendered slide images rather than objects in a PPTX file, so we now use "slide image editing" for this task in Section 1, Section 3.1, and Table 8, consistent with the title, and state this distinction in Section 3.1.

W4, Q3: Data engine and PPTX transformations

W4. Line 88: The paper itself refers to some operations as "deterministic PPTX transformations." Given this, I wonder how redundant the slide-native data engine is in practice. It would be helpful to provide statistics showing how often or to what extent this component is actually needed.

Q3. Line 88: Given that the paper itself refers to certain operations as "deterministic PPTX transformations," how redundant is the slide-native data engine in practice? Do the authors have statistics quantifying how often this component is required?

We apologize for the confusing wording. The "slide-native data engine" is our name for the entire data construction pipeline, which builds both generation and editing data. "Deterministic PPTX transformations" refers to one component of it: we construct image editing pairs by executing edits on PPTX files with code and rendering the result, so each target image is exact and reproducible, unlike methods that construct editing data by synthesizing targets with an image generator. Editing pairs make up about 16% of the continual-pretraining data (Table 7). We have revised line 88 to "code-executed PPTX edits" and clarified the terminology in Section 3.2.

W5, W6, Q4: Iterative editing

W5. In PPT slide editing, iterative editing is inevitably needed. For example, with Nano-Banana-2, repeated iterations in PowerPoint slide generation can cause text to become corrupted or image quality to gradually degrade. I wonder whether the proposed model exhibits similar behavior.

W6. I am not suggesting that the authors necessarily need to eliminate this phenomenon, but I believe such behavior, if present, should be reported as a limitation of the paper.

Q4. Does the proposed model suffer from degradation under repeated iterative editing? For example, repeated edits with Nano-Banana-2 in PowerPoint slide generation can lead to corrupted text or progressively degraded image quality. If similar behavior occurs in the proposed system, could the authors report it as a limitation?

Yes, and we now report it. We added a supplementary experiment with eight 20-round editing chains on text-dense slides from VisSlideBench-Edit, alternating a local edit (title color, one word, paragraph font, keyword bolding, or background color) with a request to undo it, each round taking the previous output as input. A round trip that changes only the title color leaves the slide essentially unchanged, but text-level edits (replacing a word, changing a paragraph's font) begin to corrupt small body text, later rounds do not repair it, and the errors accumulate; by round 20, small body text is largely garbled or missing in most chains, consistent with what the reviewer observes for Nano-Banana-2. This follows from regenerating the full slide in every round with a model trained on single-round editing pairs. We report it as a limitation in Appendix E and show a representative chain in Appendix H. In the hybrid workflow of Appendix H, text is rebuilt as native PPTX objects, so later text edits do not pass through repeated image regeneration.

W7: Section 6.5 title

W7. Section 6.5: The title "More Analysis" sounds somewhat awkward. It may be better to separate the two analyses into Sections 6.5 and 6.6, or use a title that captures what the two analyses have in common.

Thank you. We have split it into Section 6.5 (Verifier-Guided Test-Time Scaling) and Section 6.6 (Robustness to Prompt Format).

W8, W9: External evaluation

W8. The most critical weakness is the evaluation setup. Evaluating the authors' main system primarily on a benchmark created by the authors themselves weakens the credibility of the evaluation.

W9. The proposed method should also be sufficiently evaluated on the other existing benchmarks cited in the paper.

We agree that results on our own benchmark need external support. InfoVis is now evaluated on four benchmarks that we did not construct, covering single-slide generation, slide image editing, and the document-to-slide workflow. BizGenEval is reported in the paper; SlidesBench and TSBench are supplementary experiments, now reported in Section 6.3 (Table 4) and Appendix H; on SlidesGen-Bench, we re-ran the workflow with full-length decks and added the content track.

BizGenEval (single-slide generation). On the Slides domain of BizGenEval (Li et al., 2026), InfoVis reaches 46.0 / 79.8 on the hard / easy splits, the best result among the open-source models in this comparison and ahead of several proprietary systems (Table 2 of the paper).

SlidesBench (single-slide generation). SlidesBench (Ge et al., 2025) scores generated slides against reference slides (reference-based, RB) and on their own (reference-free, RF). We evaluate four image models under the same protocol on its food subset (32 slides, detailed instructions with images). InfoVis scores 68.2 overall, above Qwen-Image-Edit-2511 (64.5) and its base model FLUX.2-klein-base-9B (64.1) (Table A).

Table A. SlidesBench results. For the rows we evaluated, on the food subset (32 slides), Overall is the mean of the eight RB and RF metrics. †: all values, including Overall, as reported by Ge et al. (2025) on the full test set (195 slides), which are only indicative.

MethodExecution (%)RB ElementRB ContentRB ColorRB PositionRF TextRF ImageRF LayoutRF ColorOverall
Nano-Banana-2100.074.360.020.377.773.182.591.986.970.8
InfoVis (ours)100.074.958.220.776.865.676.287.585.668.2
Qwen-Image-Edit-2511100.073.655.621.076.856.976.275.081.264.5
FLUX.2-klein-base-9B100.073.750.921.376.852.571.981.983.864.1
AutoPresent without SlidesLib†79.067.779.710.975.945.362.754.260.945.2
AutoPresent with SlidesLib†84.184.292.218.167.247.873.258.664.755.0

TSBench (slide image editing). TSBench (Jung et al., 2026) is the slide-editing benchmark released with Talk to Your Slides; its slides and instructions are independent of our data pipeline. Its instructions are written for PPTX-editing agents, so we rewrite each one into an explicit execution instruction with an LLM (DeepSeek-V4.1-Flash), and all four image editors receive the same rewritten instructions. We evaluate the 263 instructions executable on a single slide image, and GPT-4.1 scores each output with the benchmark's prompts. InfoVis outperforms Qwen-Image-Edit-2511 and FLUX.2-klein-base-9B in total score, most clearly in text quality (4.30 vs. 3.51 and 3.77) and instruction adherence (4.12 vs. 3.44 and 3.64; Table B).

Table B. TSBench results on the 263 instructions executable on a single slide image (GPT-4.1 judge with the benchmark's prompts; each score 0–5, Total out of 30).

EditorInstruction adherenceVisual qualityText qualityImage qualityLayout qualityColor qualityTotal
Nano-Banana-24.734.744.754.954.844.9528.96
InfoVis (ours)4.124.154.304.904.644.8826.99
FLUX.2-klein-base-9B3.643.983.774.944.584.8925.80
Qwen-Image-Edit-25113.443.383.514.623.964.8323.73
Reviewer_rPedOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents InfoVis, a unified 9B model for slide image generation and editing. It is trained on a large slide-specific dataset. The authors also introduce VisSlideBench, a benchmark covering four slide-centric tasks.

Critical strengths
  1. Complete pipeline. The work covers the full chain from data construction to training to evaluation, and the problem is addressed in a way that I find reasonable and without obvious flaws.
  2. Commitment to release. The authors commit to releasing the code, model weights, and data, which increases both the value of this work to the community and its credibility.
Critical weaknesses
  1. Limited academic novelty. Overall, this is an engineering- and application-oriented paper. The data construction is fairly conventional, and the pipeline contains limited innovation. In terms of academic novelty alone, the paper did not teach me anything new. I therefore lean towards borderline reject.

I should note that I am not familiar with the current state of this specific sub-field. If reviewers more familiar with the area indicate that the data construction or the resulting model breaks some industry-standard assumption I was not aware of, or has particularly important application value to the field, I am willing to raise my score.

Questions for authors

None

3 comments

We thank the reviewer for the candid assessment and for being open to raising the score. The reviewer names two criteria: whether the data construction or the resulting model breaks an industry-standard assumption, and whether the work has particularly important application value. We address both below; all changes in the revised paper are marked in blue, and table numbers refer to the revised paper. The reviewer also invited the views of reviewers more familiar with this area; the other two reviewers' assessments speak directly to both criteria, as noted below.

W1: Novelty and application value

1. Findings that challenge common assumptions. Reviewer vAeW finds that our training-stage ablations "reveal informative generation–editing trade-offs and support the effectiveness of mixed-task training and parameter fusion." Together with a matched-budget experiment we added, these ablations yield three findings, each challenging a common default in building unified generation and editing models:

  • Domain adaptation with generation data alone does not carry over to editing. A common assumption is that adapting a unified model to a new domain with generation data also benefits its editing. Yet text-to-slide-only continual pretraining raises text-to-slide content faithfulness (CF) from 6.29 to 8.00 but drops editing semantic consistency from 5.98 to 1.60; four-task mixed training restores it (7.87; Table 5, Section 6.4).
  • High-quality synthetic SFT data can hurt real slides. SFT data rendered by a strong proprietary model is commonly assumed to help. Yet fine-tuning on our synthetic multi-image SFT data lowers CF on real-deck cases from 8.64 to 7.99 while barely changing synthetic cases (8.34 to 8.29); checkpoint fusion recovers both (8.79 and 8.61; Section 6.4, Appendix C.4).
  • Joint multi-task fine-tuning, the standard recipe for unified models, is not the best use of a fixed budget. Starting from the same shared checkpoint and at a matched budget (about 25% of the SFT compute used after that checkpoint), fusing four task-specific branches scores 7.78 overall on VisSlideBench, whereas joint four-task SFT scores 7.68, below the shared checkpoint itself (7.73). Fusion leads on 8 of the 11 metrics, most clearly on editing semantic consistency (8.74 vs. 8.30; Table A). We attribute this to the different target distributions: generation targets are rendered by Nano-Banana-2, whereas half of the editing pairs come from executed PPTX edits, so separate branches avoid fitting one set of weights to both.

Table A. Joint SFT and specialization followed by fusion at a matched budget on VisSlideBench. Both start from the shared checkpoint; Overall averages the eleven metrics, as in Table 1 of the paper. CF, IF, AQ: content faithfulness, instruction following, aesthetic quality; SC, PQ: semantic consistency, presentation quality.

ModelT2I CFT2I IFT2I AQTemplate CFTemplate IFTemplate AQMulti-image CFMulti-image IFMulti-image AQEdit SCEdit PQOverall
Shared checkpoint (starting point)8.287.806.838.257.876.818.507.826.828.447.607.73
Joint four-task SFT8.197.797.048.107.916.878.387.636.838.307.477.68
Specialization + fusion8.347.876.978.067.906.938.537.856.888.747.557.78

2. Application value. Reviewer vAeW considers the data engine and benchmark "useful resources," and Reviewer N8iN agrees that slide images differ from the distributions of general image generators and finds the image-generation perspective on slides "an interesting direction." The following points substantiate this:

  • Generating text-rich visuals, with accurate text, coherent layout, and supplied images on one canvas, is challenging for open-source models: the best open-source baseline (FLUX.2-dev) scored 6.37 overall on VisSlideBench; InfoVis reaches 7.94, against 7.99 for Nano-Banana-2. We will release the model, training data, data engine, and VisSlideBench to support the open-source community in this direction.
  • Image-based slide generation is already deployed: SlidesGen-Bench (Yang et al., 2026) classifies NotebookLM, Skywork-Banana, and Kimi-Banana as image-generation systems, showing that slide generation is a practical use case for image generation models.
  • In a supplementary experiment (Appendix H), InfoVis serves as the visual layer of a Codex-based pipeline that produces an editable PPTX, showing that image generation models can be integrated into current agentic slide-creation workflows and enrich the diversity and aesthetic quality of the images embedded in the slides.

Dear Reviewer rPed, as the discussion period is coming to an end, we would like to briefly follow up. You noted that you would be willing to raise your score if reviewers more familiar with the area found that the work breaks an industry-standard assumption or has particularly important application value. Reviewer vAeW, with a confidence of 4 ("I have substantial direct research experience on this topic"), recommends weak acceptance, describes the data engine and benchmark as "useful resources," and finds that the training-stage ablations "reveal informative generation–editing trade-offs." Our response above details the three findings that challenge common practice in building unified generation and editing models, and the application value of InfoVis. We hope this meets the criteria you set out, and we would be grateful if you would reconsider your assessment in light of it. Thank you again for your time and candid feedback.

Dear Reviewer rPed, as today is the last day of the discussion period, we are sending a brief reminder in case our earlier messages were missed. We would greatly appreciate it if you could take a look at our response and follow-up above, and we would be happy to answer any remaining questions before the discussion closes. Thank you for your time.

Reviewer_vAeWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents InfoVis, a unified 9B model for text-to-slide generation, template-conditioned generation, multi-image grounded generation, and instruction-based slide editing. The work introduces a slide-native data engine containing over 13M continual-pretraining records and 1.9M supervised fine-tuning records, together with VisSlideBench, a benchmark of 1,527 instances. The model adapts FLUX.2-klein-base-9B through staged training, task-specific fine-tuning, and parameter fusion.

Critical strengths
  1. The slide-native data engine and benchmark provide useful resources covering four complementary generation and editing tasks, with human validation supporting the evaluation.
  2. InfoVis improves the starting backbone from 6.30 to 7.94 and leads the evaluated open-source baselines on all eleven metrics. External BizGenEval results further support its effectiveness.
  3. The training-stage ablations reveal informative generation–editing trade-offs and support the effectiveness of mixed-task training and parameter fusion.
Critical weaknesses
  1. The fusion ablation does not fully isolate the benefit of the training strategy. The shared SFT checkpoint receives text-to-slide SFT, whereas the final model incorporates four task-specific SFT branches. Consequently, the improvement from 7.73 to 7.94 combines additional task supervision with parameter fusion. A joint four-task SFT baseline using the same data and a comparable update budget would clarify the specific advantage of specialization followed by fusion.
  2. External validation covers a narrower scope than the four-task benchmark. As acknowledged in the paper, editing evaluation is in-distribution, and the document-to-slide workflow is evaluated externally only on aesthetics. These results support generation quality and visual presentation, while editing generalization and document-level content preservation remain less established.
Questions for authors

Refer to weaknesses.

2 comments

We thank the reviewer for the positive assessment and for the constructive suggestions on the fusion ablation and on external validation. We address each point below; all changes in the revised paper are marked in blue, with the new baseline summarized in Section 6.4 and the new external results in Section 6.3, both detailed in the new Appendix H. Table numbers refer to the revised paper.

W1: Joint SFT vs. specialization and fusion

We added a supplementary experiment with the suggested baseline: joint four-task SFT from the shared SFT checkpoint on the union of the four task datasets. Within the compute available during the rebuttal period, we compare the two strategies at a matched budget of about 25% of the SFT compute that the final model uses after the shared checkpoint; the fused model combines four task branches trained with the same total budget. On VisSlideBench, specialization followed by fusion scores 7.78 overall, against 7.68 for joint SFT (Table A). Fusion leads on 8 of the 11 metrics, with the largest margins on editing semantic consistency (8.74 vs. 8.30) and multi-image instruction following (7.85 vs. 7.63); the other three metrics differ by at most 0.07. Joint SFT also falls below the shared checkpoint it starts from (7.73), reflecting the cross-task interference discussed in Section 6.4, whereas fusion improves on it. We attribute this to the different target distributions of the tasks. All generation SFT targets are rendered by Nano-Banana-2, whereas half of the editing SFT pairs (about 70K of 140K; Table 6) are produced by executing edits on PPTX files and rendering the result, so their targets follow the rendering of the source slide. Joint training pulls one set of weights toward both distributions at once; training each branch separately lets it fit its own target distribution, and fusion then combines the branches in weight space. The largest gap, on editing semantic consistency, is consistent with this: editing makes up only about 7% of the SFT data and is the task most easily pulled toward the generation distribution. Specialization followed by fusion is therefore more effective than joint training at the same update budget, and it keeps the tasks modular, since each branch can be trained or updated independently.

Table A. Joint SFT and specialization followed by fusion at a matched budget on VisSlideBench. Both start from the shared checkpoint; Overall averages the eleven metrics, as in Table 1 of the paper.

ModelT2I CFT2I IFT2I AQTemplate CFTemplate IFTemplate AQMulti-image CFMulti-image IFMulti-image AQEdit SCEdit PQOverall
Shared checkpoint (starting point)8.287.806.838.257.876.818.507.826.828.447.607.73
Joint four-task SFT8.197.797.048.107.916.878.387.636.838.307.477.68
Specialization + fusion8.347.876.978.067.906.938.537.856.888.747.557.78

W2: External validation

We added supplementary experiments for both points, and both show that the strengths of InfoVis carry over beyond our own benchmark. On TSBench, an external slide-editing benchmark, InfoVis outperforms Qwen-Image-Edit-2511 and FLUX.2-klein-base-9B. On SlidesGen-Bench, our document-to-slide workflow, re-run with full-length decks, achieves the second-highest overall content accuracy among the systems that support every task type (87.72%, behind only Zhipu) and the highest aesthetics score among the ten systems (27.95). Details follow.

  • Editing generalization. TSBench (Jung et al., 2026) is the slide-editing benchmark released with Talk to Your Slides. Its instructions are written for PPTX-editing agents, so an LLM rewrites each into an explicit execution instruction, which all editors share. We evaluate the 263 instructions that can be executed on a single slide image, and GPT-4.1 scores each output with the benchmark's prompts on instruction adherence and five quality dimensions (0–5 each). InfoVis outperforms its base model FLUX.2-klein-base-9B and Qwen-Image-Edit-2511 in total score (26.99 vs. 25.80 and 23.73), most clearly in text quality (4.30 vs. 3.77 and 3.51) and instruction adherence (4.12 vs. 3.64 and 3.44; Table B). Since TSBench's slides and instructions were built independently of our data pipeline, this shows that the editing ability of InfoVis generalizes beyond its training distribution.
  • Document-level content preservation. SlidesGen-Bench (Yang et al., 2026) evaluates content with QuizBank: for each source document it builds ten multiple-choice questions, and an LLM answers them using only the content extracted from the generated slides. We re-ran our workflow on all 189 documents with full-length decks (15 slides per document) and now report this content track. InfoVis reaches an overall QuizBank accuracy of 87.72%, the second highest among the systems that support every task type, behind only Zhipu (88.35%); it is also above the average of the reported systems in every category and difficulty column and the highest of all reported systems on business plans, product launches, and high-difficulty questions (Table C).

Table B. TSBench results of the open-source editors on the 263 single-slide instructions (GPT-4.1 judge; each score 0–5, Total out of 30).

EditorInstruction adherenceVisual qualityText qualityImage qualityLayout qualityColor qualityTotal
InfoVis4.124.154.304.904.644.8826.99
FLUX.2-klein-base-9B3.643.983.774.944.584.8925.80
Qwen-Image-Edit-25113.443.383.514.623.964.8323.73

Table C. QuizBank accuracy (%) on SlidesGen-Bench. Except for InfoVis, per-column values (including the average of the eleven reported systems) are from Table 1 of Yang et al. (2026), which has no overall column. Overall is the mean of the three difficulty columns, which equals the accuracy over all 189 documents because the difficulty levels are balanced; it is computed only for systems that support every task type (–: not computed). N/A: task type not supported; *: pipelines adapted by Yang et al. (2026). Bold: highest in each column.

SystemOverallBrand promotionBusiness planPersonal statementProduct launchCourse preparationWork reportTopic introductionHigh difficultyLow difficultyMedium difficulty
InfoVis + workflow (ours)87.7286.6782.5093.0092.0888.6786.1586.1784.6090.3288.25
Average of the 11 systems–76.1957.9385.3075.3181.4475.3776.3972.7782.7075.84
Gamma70.4263.5752.5074.0054.4485.0048.4676.3467.2675.5668.45
Kimi-Banana87.1984.2977.5096.0090.4384.0085.3886.2283.3391.5886.67
Kimi-Smart–N/AN/A95.00N/AN/A84.2979.6878.7888.5078.57
Kimi-Standard–N/AN/A96.47N/AN/A80.0076.8875.7188.0575.43
NotebookLM–83.0045.00N/A84.2193.0086.1568.3369.8186.0071.16
Quark–84.2951.2594.0079.05N/A68.0082.1369.7890.5282.12
Skywork81.3080.0068.7578.5076.5284.6786.1583.1280.0082.5881.31
Skywork-Banana83.9992.1467.5091.5089.5787.6990.0079.3379.6788.0084.31
Zhipu88.3580.0068.7596.0086.0890.0084.1790.4484.0792.0088.97
PPTAgent*62.4662.2547.3769.1561.8066.8661.2761.9458.9466.7661.68
AutoPresent*56.3256.1342.7262.3555.7360.2955.2555.8553.1560.2055.62

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We thank all reviewers for their constructive feedback. Reviewer vAeW considers our data engine and benchmark "useful resources"; Reviewer rPed finds that the work "covers the full chain from data construction to training to evaluation" and considers our approach "reasonable and without obvious flaws"; and Reviewer N8iN finds the image-generation perspective on slides "an interesting direction." Below we restate our contributions and their value and summarize the supplementary experiments. All changes in the revised paper are marked in blue, the new experiments are detailed in the new Appendix H, and table numbers refer to the revised paper.

1. Contributions.

  • Data. A slide-native data engine that produces over 13M continual-pretraining and 1.9M supervised fine-tuning records for four tasks: text-to-slide, template-conditioned, and multi-image grounded generation, and slide image editing.
  • Benchmark. VisSlideBench, 1,527 manually screened instances that evaluate all four tasks with task-specific, multidimensional metrics validated against expert rankings.
  • Model. InfoVis, a unified 9B model that leads every evaluated open-source model on all eleven VisSlideBench metrics and scores 7.94 overall, against 7.99 for Nano-Banana-2.

2. Empirical findings. Our training analysis yields three empirical findings on building unified generation and editing models:

  • Adapting a model to slides with generation data alone erodes editing. Text-to-slide-only continual pretraining drops editing semantic consistency from 5.98 to 1.60; four-task mixed training restores it.
  • Synthetic SFT data from a strong proprietary model can hurt real slides. Synthetic multi-image SFT lowers content faithfulness on real-deck cases from 8.64 to 7.99; checkpoint fusion recovers it.
  • Specialization followed by fusion makes better use of a fixed budget than joint multi-task fine-tuning. At a matched budget, it scores 7.78, against 7.68 for joint SFT.

3. Application value: image generation and programmatic generation are complementary. Slide generation is already a practical use of image generation: SlidesGen-Bench classifies NotebookLM, Skywork-Banana, and Kimi-Banana as image-generation systems. Yet open-source models lagged far behind: the best open-source baseline scored 6.37 overall on VisSlideBench. Image generation and code-generated PPTX are not competing routes. Code gives exact text and native, editable objects, whereas image generation gives richer visual design; SlidesBench shows this pattern, with AutoPresent, which writes code, scoring higher on reference content similarity and the image models scoring higher on reference-free layout. The two routes can also be combined. In Appendix H, Codex builds an editable PPTX deck in which InfoVis generates the slide visuals and all text is rebuilt as native text boxes; when InfoVis works with such a coding agent, its slides can be edited at the object level and animated like any native deck. We will release the model, data, data engine, and benchmark to support this direction.

4. Supplementary experiments for the main concerns.

  • Evaluation beyond our own benchmark (vAeW, N8iN). InfoVis is now evaluated on four external benchmarks covering single-slide generation, slide image editing, and the document-to-slide workflow (Section 6.3, new Table 4, Appendix H). It is the best open-source model in the BizGenEval Slides comparison, and on SlidesBench and the slide-editing benchmark TSBench it outperforms Qwen-Image-Edit-2511 and its base model FLUX.2-klein-base-9B. On SlidesGen-Bench, a re-run with full-length decks obtains the highest aesthetics score among ten systems (27.95) and, on the newly added content track, QuizBank, 87.72% accuracy, second only to Zhipu among the systems that support every task type.
  • Training strategy (vAeW). A joint four-task SFT baseline from the same shared checkpoint at a matched budget isolates the benefit of specialization followed by fusion (the third finding above; Section 6.4, Appendix H).
  • Combining the two routes (rPed, N8iN). The Codex + InfoVis case study described above (Appendix H).

We hope these results address the reviewers' concerns, and we welcome further discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.