acceptodds
Withdrawn from ICLR 2027

DisciplineGen-1M: A Large-Scale Dataset for Multidisciplinary Visual Generation and Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent image generation and editing models can produce visually appealing natural images, yet they remain unreliable when the target image is a knowledge-intensive diagram whose correctness depends on disciplinary concepts, symbolic structure, and precise spatial relations. We introduce DisciplineGen-1M, a million-scale multidisciplinary dataset that supports text-to-image generation and image editing. It contains 1.2M samples spanning mathematics, physics, chemistry, biology, geography, computer science, economics, history, music, and sports. To construct the dataset, we design a scalable framework that combines structured rendering, OCR-based editing, specialized programmatic synthesis, and large-scale text-to-image filtering. These pipelines produce captions, editing instructions, structured annotations, and paired images with controllable semantic differences. Building on DisciplineGen-1M, we further introduce a discipline-informed reasoning-generation model for both text-to-image generation and image editing. Experiments on discipline-related benchmarks, GenExam and GRADE, show substantial improvements over open-source baselines, while evaluations on general reasoning-informed benchmarks, WISE and RISE, further indicate broader transfer. The results suggest that large-scale structured academic visual data is a key ingredient for moving image generation from aesthetic plausibility toward verifiable knowledge-grounded visual creation. We will publicly release our dataset, model, and source code of the data curation pipeline to ensure reproducibility and benefit future research.

Primary area:
datasets and benchmarks
Keywords:
image generation, image editing, multidisciplinary, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NHJ5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The submitted paper builds a dataset for generating and editing academic diagrams, such as geometry figures, molecules, sheet music, and state machines. It has about 1.2M samples across ten subjects. Most of the data is synthetic. The authors render it from code (SVG, TikZ, RDKit, LilyPond, etc.), mask text labels to create editing pairs, and filter open-source educational images. For the model, they use a frozen Qwen3-VL-8B to rewrite prompts. They then LoRA-finetune Qwen-Image for generation and Qwen-Image-Edit for editing. The finetuned models outperform other open-source models on GenExam (51.4 vs. 35.8 base) and GRADE (58.7 vs. 30.0 base). They also show improvements on WISE and RISE. An ablation study adds each data pipeline step by step to show its contribution.

Critical strengths
  1. Most edit pairs are made by changing the source code and re-rendering, so the two images only differ where the edit is. Clean pairs like this are hard to collect, and the data would be useful if released.

  2. Over the same base models, GenExam goes up by 15.6 and GRADE by 28.7.

  3. The data covers subjects like music and sports that prior datasets rarely include. The appendix is also detailed enough to reproduce the pipeline.

Critical weaknesses
  1. It is unclear how much of the gain comes from the data itself. Prompt enhancement alone gives a large part of the improvement, and the ablation does not control training budget.
  2. The dataset is not well documented or verified. The claimed 1.2M does not match the training size in Appendix C, and there is no quantified human audit.
  3. The benchmark gains are clear, and the WISE and RISE results provide some evidence of transfer. However, the paper does not report controlled evaluations on unseen templates or edit types, and test-set overlap checks are not fully described for the synthetic pipelines.
Questions for authors

Q1: Effect of prompt enhancement. Table 6 shows prompt enhancement alone brings large gains (GenExam 35.8→41.0, GRADE 30.0→41.2, WISE 0.50→0.59, RISE 19.4→23.5). There is no result for the finetuned model without it, and the baselines in Tables 2–5 do not use the same rewriter. Can you report the full 2×2 setting (finetuned or not × prompt enhancement or not), and run the main baselines with the same rewriter?

Q2: Ablation design. (a)In Table 6, each data type is added in a fixed order, and every step also adds more samples and training steps. The gains could come from having more data rather than from the data type itself, and a different order might give a different picture. (b)I am also confused by two rows in Table 6: OCR editing data is added to the T2I model, and T2I filtering data is added to the editing model. Section 4.1 says each model only uses its own task data, so how were these used? (c) It would help to see leave-one-out results with the total training budget fixed. (d)A comparison with an existing dataset such as StructVisuals or S1-Omni-Image, using the same base model and budget, would also show whether this data is better than what is already available.

Q3: The training setup in Appendix C (43,993 and 23,384 steps, effective batch size 8, one pass, repeat factor 1) implies about 352K T2I and 187K editing samples, about 539K in total. How does this relate to the reported 1.2M? Could you give the sample counts by pipeline before filtering, after filtering, and used in training?

Q4: Appendix B.3 mentions manual checks, but only for the SVG branch built from textbook knowledge (economics and biology), and no numbers are given. The other pipelines seem to rely only on compilers, rules, or VLM judges. Could you report a human audit on a sample from each pipeline and subject, showing how often there are knowledge errors, mismatches between the instruction and the target, or changes outside the edit region?

Q5: Some training tasks are similar to those evaluated in the benchmarks, such as the Go and music cases in Figure 5. This is reasonable for coverage, but generalization to unseen task types remains unclear. Could you evaluate held-out templates or edit types? Appendix B.5 describes pHash-based benchmark overlap checks for retrieved images. Were synthetic samples also checked at the source level, such as shared game positions, score fragments, or graph instances?

Minor:

  1. Table 4 incorrectly lists SenseNova-U1-8B-MoT as closed-source, although its weights are publicly available. Its reported score is also higher than the proposed model’s (0.68 vs. 0.66). Please correct the classification and any affected ranking statements.
  2. Section 4.3 claims gains over the strongest closed-source baseline on RISE, but Table 5 reports 26.4 for the proposed model versus 49.4 for GPT-Image-2. Please clarify or correct this statement.
Reviewer_MAPHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces DisciplineGen-1M, a 1.2M-sample dataset spanning ten disciplines and supporting both text-to-image generation and image editing through four construction routes. The accompanying two-stage system uses a frozen Qwen3-VL-8B module to expand implicit requests into explicit plans, followed by task-specific Qwen-Image generators adapted with LoRA for generation or editing.

Critical strengths

The paper introduces DisciplineGen-1M, a large-scale dataset containing 1.2M samples across ten disciplines, covering both text-to-image generation and image editing.

Critical weaknesses
  1. Lack of systematic dataset quality evaluation. Dataset quality is central to the paper's contribution, yet the evaluation primarily focuses on downstream model performance. Although the authors describe automated filtering and VLM-based quality checks, they provide no quantitative human evaluation. The reliability of the VLM-based quality judges is also not validated. Given the knowledge-intensive nature of the dataset, these omissions make it difficult to assess how trustworthy the 1.2M samples actually are.

  2. Insufficiently controlled comparisons and unclear attribution of performance gains. The proposed approach combines dataset-based fine-tuning with reasoning-based prompt enhancement, but it is unclear whether the baseline models receive the same enhanced prompts. Without evaluating the fine-tuned models under the original prompts, the contribution of DisciplineGen-1M cannot be clearly isolated. Furthermore, the cumulative data ablations do not control for training data size, making it difficult to distinguish the benefits of individual construction pipelines from the effect of additional training data. The paper also lacks comparisons with existing datasets under matched model architectures and training budgets.

Questions for authors

Have the authors conducted any human evaluation of the final dataset? How reliable are the VLM-based quality assessments? Can the authors report their agreement with human experts on a representative sample?

Were the same reasoning-based prompt enhancements applied to all baseline models in Tables 2–5? If not, how do the baselines perform with the same enhanced prompts?

How do the fine-tuned models perform without prompt enhancement?

Can the authors provide fixed-data-size ablations or comparisons with existing structured visual datasets under matched training settings?

Reviewer_FA53OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a dataset for infographic generation, proposing a series of data construction pipelines and ablation results. and gives some practical solutions for getting a good performance.

Critical strengths
  1. Infographic generation and editing are important topics; I recognize the contribution for construction a large-scale dataset.
  2. Some of the data construction pipelines are useful, especially the ones using coding tools.
  3. The authors provide plenty of experimental results.
Critical weaknesses
  1. The biggest concern is that this paper lacks novelty. I recognize the "contribution" of collecting data, but I do not see anything that other people could not do. I think every company is able to collect data using these "relatively regular" pipelines.
  2. Although some of the construction pipelines are useful, the most challenging part of infographic data construction is not touched. The most challenging part is how to give a "precise" caption for complicated graphs. Most of the VLMs would have many hallucinations.
  3. The experimental results are predictable; no new information is provided.
  4. Prompt enhancement is critical, but this is challenging for infographics. How to push the boundary of PE is important. But is not touched in this paper.
Questions for authors

Please refer to the weaknesses. I think the most important parts are the caption(for training) and PE (for inference). Please provide some insight and exploration. Please summarize the "new things" in this paper, if there are any.

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

Related papers

Discussion (0)

Sign in to comment.