PaperlayerView exact revision on arXiv

Source-linked AI summary

SciForma: Structure-Faithful Generation of Scientific Diagrams

Yuxuan Luo, Peng Zhang, Xinjie Zhang, Xun Guo, Zhouhui Lian, Yan Lu

arXiv:2607.18091v1cs.CVcs.GRcs.LG

TL;DR

Scientific methodology diagrams require correct components, directional relations, and text, but current approaches lack reliable per-axis verification and enforcement. SciForma combines a three-axis structural inventory with M-DPO and iterative editing, and SciForma-9B exceeds open-source baselines and GPT-Image-1.5 on SciFormaBench-2K and AIBench.

  • Problem

    Scientific methodology diagrams require legible components, directional relations, and annotations, yet existing open-source approaches lack reliable per-axis verification and enforcement of structural correctness.

    Source: §1 · p. 2

  • Method

    SciForma uses a Component–Arrow–Text structural inventory, curated training and evaluation datasets, M-DPO for conjunctive correctness with adaptive gradient routing, and verification-gated iterative editing.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    SciForma-9B exceeds all open-source baselines and GPT-Image-1.5 on SciFormaBench-2K and AIBench, scoring 69.51 on SciFormaBench-2K.

    Source: §1 · p. 2, §5.2 · p. 6

  • Takeaways & Limitations

    SciForma-9B brings open scientific diagram generation close to proprietary-level structural fidelity and aesthetics with explicit structural constraints.

    Source: §7 · p. 8

    SciForma’s 1024 px resolution limits dense-diagram detail, its evaluation-and-editing pipeline relies on a proprietary VLM, and agentic inventories may drift from user intent.

    Source: §6 · p. 8

Abstract

from arXiv · show

Structural fidelity is essential to scientific methodology diagrams. To communicate research logic, these diagrams must faithfully render components, directional relations, and textual annotations. Since a single error, such as a reversed arrow or an unreadable equation, can invalidate the entire figure, structural fidelity is inherently conjunctive: correctness on one axis cannot compensate for failure on another. Current open-source models fail to satisfy this criterion. Supervised fine-tuning (SFT) learns plausible layouts but cannot reliably ensure structural correctness, while scalar reward-based post-training obscures which structural dimension has failed. To address this, we introduce SciForma, a framework for the structure faithful generation of scientific methodology diagrams. Specifically, SciForma decomposes diagram quality into three structural axes: Component, Arrow, and Text, guided by a structural inventory. Built on this foundation, we curate SciFormaData-700K for structured training and SciFormaBench-2K for logic-verified evaluation. To close the gap left by SFT, we develop Multi-Dimensional Conjunctive Preference Optimization (M-DPO), which enforces simultaneous correctness across all axes and adaptively routes gradients to the most deficient dimension in post-training. The same structural inventory also enables iterative editing at inference time to correct residual errors. This combination allows SciForma-9B to exceed all open-source baselines and GPT-Image-1.5 on both SciFormaBench-2K and AIBench, bringing open scientific diagram generation close to proprietary-level structural fidelity. Our code and data will be available at: https://github.com/microsoft/SciForma.

1 Introduction

SciForma treats scientific-diagram fidelity as conjunctive correctness across components, arrows, and text, and uses structural verification to guide training and refinement. SciForma-9B surpasses open-source baselines and GPT-Image-1.5 on SciFormaBench-2K.

1 Introduction

  • Current SFT can learn plausible layouts without guaranteeing structural correctness, while scalar post-training signals obscure which structural dimension failed.The introduction attributes the open-source fidelity gap versus proprietary systems [Google DeepMind 2025; OpenAI 2026b] to the lack of mechanisms for independently verifying and enforcing each axis; existing post-training methods include [Rafailov et al. 2023; Shao et al. 2024; Wang et al. 2026].Source: §1 · p. 2
  • SciForma decomposes diagram quality into independently verifiable Components, Arrows, and Text, using a structural inventory to extract ground-truth checklists from reference diagrams.This reframes evaluation as explicit structural verification rather than continuous visual matching, so strengths on one axis cannot obscure failures on another.Source: §1 · p. 2
  • SciForma combines SciFormaData-700K and SciFormaBench-2K with M-DPO, which contrasts a global winner against axis-specific losers and routes gradients toward the most deficient axis.At inference, the structural inventory supports iterative, verification-gated inpainting to identify and correct residual errors.Source: §1 · p. 2
  • SciForma-9B exceeds all open-source baselines and GPT-Image-1.5 on SciFormaBench-2K, with M-DPO’s largest gains on the Arrow axis and Hard tier.On AIBench, it exceeds human-drawn originals with the largest margin on Topology.Source: §1 · p. 2
  • On AIBench, SciForma-9B exceeds human-drawn originals, with its largest margin on Topology, and generalizes to agentic-generated prompts in PaperBanana.Ablations show that M-DPO breaks SFT stagnation, avoids scalar-reward collapse, and needs decomposed pairs coupled with a conjunctive objective for consistent cross-axis gains.Source: §1 · p. 2

2 Related Work

Prior scientific-diagram generation methods leave structural-fidelity gaps, and existing alignment methods do not enforce conjunctive correctness across independently verifiable structural axes.

2.1 Scientific Diagram Generation

2.2 Post-training Alignment for Generative Models

  • Existing alignment methods reduce per-axis feedback to scalar objectives, and none had addressed independently verifiable structural axes whose single-axis failure invalidates output; M-DPO targets that gap.Recent multi-dimensional preference methods address perceptual axes that co-vary smoothly, while Neighbor GRPO still collapses feedback into a scalar [He et al. 2026].Source: §2.2 · pp. 2–3

3 Structural Inventory

SciForma formalizes structural fidelity through three independently verifiable primitives and evaluates outputs against a checklist extracted from the prompt and reference image. Per-axis scores make discrepancies diagnosable.

3.1 Universal Structural Primitives

  • UML activity diagrams [Object Management Group 2017] and BPMN process models [Object Management Group 2011] exemplify a shared grammar of typed entities, directed relations, and textual qualifiers.This grammar motivates the inventory’s component, arrow, and text primitives.Source: §3.1 · p. 3
  • The inventory defines Component, Arrow, and Text as independently verifiable primitives for entity layout, information topology, and textual annotations.Together, they form a ground-truth checklist of what a structurally faithful diagram must contain.Source: §3.1 · p. 3

3.2 Automated Inventory Extraction and Evaluation

  • A VLM extracts a JSON checklist from the prompt and reference image, recording component coordinates, arrow source-target pairs, and literal text strings.The checklist specifies the required content for structural fidelity.Source: §3.2 · p. 3
  • Discrepancies are weighted critical (w_e=1.0) when missing or fundamentally wrong, and moderate (w_e=0.5) when present but defective; per-axis scores remain primary for diagnosis, rewards, and ablations.The per-axis average provides a compact summary, but the individual axis scores are the basis for conjunctive correctness assessment.Source: §3.2 · p. 3

4 SciForma

SciForma combines data curation, supervised fine-tuning, multi-dimensional preference optimization, and iterative refinement to address structural errors in scientific methodology diagrams.

4.1 Overview

  • Figure 2 presents SciForma’s pipeline as data curation, SFT, M-DPO, and iterative refinement.Source: §4.1 · p. 3

4.2 SciFormaData-700K

  • SciFormaData-700K pairs methodology diagrams with Component, Arrow, and Text captions and includes generation pairs and localized editing triplets.The collection draws 726K diagrams from 593K arXiv LaTeX sources, using dual-VLM consensus filtering and pHash deduplication; its 70K edits are axis-specific, verified, and focus on arrows (54%) and shapes (46%).Source: §4.2 · pp. 3–4, Figure 3 · p. 4

4.3 SciFormaBench-2K

  • SciFormaBench-2K evaluates structural fidelity with element-level attribution and axis-level decomposition, addressing failure modes obscured by holistic scores or preference-based benchmarks.Its 2,000 candidates are complexity-stratified and paired with human-verified inventories; 56.8% require formula rendering.Source: §4.3 · p. 4

4.4 Supervised Fine-Tuning (SFT)

  • Two-stage SFT adapts the base model to diagram structure and joint generation-editing, producing SciForma-Base but leaving residual errors for M-DPO.Stage 1 uses the 656K training set at about 768 px; Stage 2 co-trains generation samples and editing triplets at about 1024 px, supervising target tokens for edits.Source: §4.4 · p. 4

4.5 Multi-Dimensional Conjunctive Preference Optimization

  • M-DPO addresses the limits of scalar preference optimization by preserving axis-specific signals and enforcing simultaneous correctness across structural dimensions.Its design combines dimension-anchored preference pairs with a conjunctive preference objective; a single error can invalidate a figure.Source: §4.5 · p. 4

4.5.1 Dimension-Anchored Preference Construction.

  • For each prompt, M-DPO selects a shared top candidate as winner and axis-specific losers that are weak on one dimension but competitive on others.Axes without candidate variance are excluded from that prompt’s loss, and sharing one winner avoids gradient conflicts.Source: §4.5.1 · p. 5

4.5.2 Conjunctive Preference Objective.

  • M-DPO’s conjunctive objective requires the winner to beat every axis-specific loser, concentrating gradients on deficient axes and vanishing only when all axes are satisfied.It is derived from a multi-way Bradley–Terry model; when D=1, it reduces to standard DPO.Source: §4.5.2 · p. 5

4.6 Iterative Refinement

  • SciForma-9B refines dense diagrams through localized edits accepted only when they improve local scores and preserve or improve global Component, Arrow, and Text consistency.GPT-5.4 and a structural inventory localize and rank defects; a global-score drop triggers rollback to the best known state.Source: §4.6 · p. 5

5 Experiments

Experiments evaluate SciForma on SciFormaBench-2K, AIBench, and PaperBanana, and use ablations to test M-DPO’s structural objectives. Results report gains from M-DPO and iterative refinement across benchmark performance and human evaluation.

5.1 Implementation Details

  • Implementation uses two-stage SFT followed by M-DPO preference training, with EMA checkpoints and structural-inventory scoring.SFT uses 656K generation pairs followed by 244K high-quality pairs and 70K editing triplets; M-DPO rolls out 12 candidates for each of 50K prompts and trains on approximately 16K valid tuples.Source: §5.1 · p. 5

5.2 Results on SciFormaBench-2K

  • SciForma-9B reaches 69.51 overall on SciFormaBench-2K, surpassing GPT-Image-1.5 [OpenAI 2026a] (68.96) after SciForma-Base lifts FLUX.2-klein-base-9B [Labs 2025] from 33.87 to 67.59; iterative editing adds 2.89 to reach 72.40.The largest M-DPO gains are on Arrow (+1.82) and Text (+3.16); Arrow again gains most with iterative refinement.Source: §5.2 · p. 6, Table 2 · p. 6

5.3 Results on AIBench

  • SciForma-9B scores 70.29 on AIBench, edging human-drawn originals (70.09) and GPT-Image-1.5 [OpenAI 2026a] by 8.67 points, with Topology +6.19 over originals.Originals lead on Component (82.65 vs. 77.53); the authors suggest dense diagrams at 1024 px are harder to render clearly. AIBench’s VQA evaluation is independent of SciForma’s structural inventory.Source: §5.3 · p. 6, Table 2 · p. 6

5.4 Agentic Integration on PaperBanana

  • On PaperBanana [Zhu et al. 2026b], SciForma achieves 30.7% overall, surpassing GPT-Image-1.5 [OpenAI 2026a] and nearing Nano Banana Pro on Faithfulness (34.5% vs. 45.8%), but trails on Conciseness (37.7% vs. 80.7%).The evaluation uses 292 samples, a prompt rewriter, best-of-3 selection, and Gemini-3-Pro pairwise judgments against human-drawn originals.Source: §5.4 · p. 6, Table 3 · p. 6

5.5 Ablation Studies

  • M-DPO gains +1.92 versus continued SFT’s +0.37 in 4K versus 30K steps, improving Text +3.16 and Arrow +1.82; GDRO shuffles errors and GRPO worsens overall.The compared runs start from the same 90K SFT checkpoint; scalar rewards can hide regressions on individual axes.Source: §5.5 · pp. 6–7, Figure 5 · p. 7, Table 4 · p. 7
  • Conjunctive M-DPO gains +1.92 versus Scalar DPO’s +0.83; Pareto DPO regresses Text, mean DPO shrinks the learning rate, and CaPO/MCDPO regress Text (−1.34/−1.85), unlike M-DPO’s +3.16.The ablations attribute the gains to conjunctive coupling around a shared winner, which concentrates gradients on the weakest axis; CaPO and MCDPO are multi-dimensional DPO variants.Source: §5.5 · p. 7, Table 5 · p. 7, Table 6 · p. 7

5.6 Qualitative Results

  • Qualitative examples show accurate text and spatial layouts; checklist-scored iterative edits correct shapes, arrows, and text with rollback, while M-DPO yields cleaner, more coherent topologies.When an edit lowers global consistency, the system reverts to the previous state.Source: §5.6 · pp. 7–8

5.7 User Study

  • In a study of 36 AI graduate students judging 30 samples, SciForma-9B outperforms Wan2.7-Image across metrics and rivals Nano Banana Pro; benchmark scores correlate with preferences (r=0.76).Participants evaluated Structure, Text, Visual Quality, and Overall Preference; the reported correlation is Pearson’s per-sample correlation between SciFormaBench-2K scores and user preferences.Source: §5.7 · p. 8, Figure 6 · p. 8

6 Discussion and Limitations

SciForma’s discussion identifies limits in rendering resolution, evaluation, and agentic intent alignment, and calls for community extensions alongside human oversight to reduce misuse.

6 Discussion and Limitations

  • The 1024 px resolution, proprietary VLM evaluator, and unverified agentic inventories limit SciForma’s rendering, evaluation, and alignment with user intent.The authors identify higher-resolution training, open-source VLM evaluators, and human-in-the-loop agentic frameworks as next opportunities.Source: §6 · p. 8
  • SciForma is released to advance open image-based diagram research, with an invitation to extend M-DPO’s conjunctive design to per-axis tasks.Source: §6 · p. 8
  • To prevent misuse in paper fabrication, the authors advocate human verification and in-the-loop editing of all generated outputs.Source: §6 · p. 8

7 Conclusion

SciForma defines diagram quality through three verifiable axes—Component, Arrow, and Text—and uses M-DPO to enforce their joint correctness. Experiments show SciForma-9B approaches proprietary-level structural fidelity and aesthetics with explicit structural constraints.

7 Conclusion

  • SciForma formulates diagram quality across Component, Arrow, and Text, and uses M-DPO to enforce their joint correctness.The framework treats all three axes as verifiable structural criteria.Source: §7 · p. 8
  • Experiments show SciForma-9B approaches proprietary-level structural fidelity and aesthetics with explicit structural constraints.Source: §7 · p. 8

Supplementary Material Overview … A.1.4 Stage 4: Editing Triplet Construction.

The supplement covers dataset and benchmark details, theoretical analysis, implementation, ablations, qualitative results, and user-study materials. It also describes the four-stage SciFormaData-700K construction pipeline.

Supplementary Material Overview

  • Appendices A–I cover SciFormaData-700K, SciFormaBench-2K, M-DPO analysis, implementation, further analyses, ablations, qualitative results, user studies, and prompt templates.Source: Supplementary Material Overview · p. 11

A.1 Dataset Construction

  • SciFormaData-700K is assembled through a four-stage pipeline, with VLM-mediated prompt templates provided in Appendix I.Source: §A.1 · p. 11

A.1.1 Stage 1: Source Extraction.

  • The pipeline parses LaTeX from 593K arXiv papers across 17 cs.* categories, extracts figures and references, and resolves custom macros into structured metadata.Direct LaTeX extraction avoids PDF-parsing artifacts.Source: §A.1.1 · p. 11

A.1.2 Stage 2: Consensus Filtering.

  • About 1.8M candidates are filtered for methodology diagrams using dual-VLM consensus, while pHash removes structurally redundant diagrams and reduces overlap between training and benchmark sets.Source: §A.1.2 · p. 11
  • Qwen3-VL generates Component, Arrow, and Text captions; complexity stratification yields a balanced 2,000-sample benchmark and 656K generation training pairs.The complexity tiers are Low (13.6%), Medium (48.2%), and High (38.2%).Source: §A.1.3 · p. 11

A.1.3 Stage 3: Captioning and Stratification.

  • From a ∼252K high-quality subset, SAM3 [Carion et al. 2025] supports validated editing triplets, producing 70K high-purity edit pairs for Stage-2 SFT.The pipeline samples 1–3 components or arrows for deterministic edits and discards invalid branches through traceback.Source: §A.1.4 · p. 11

A.1.4 Stage 4: Editing Triplet Construction.

A.2 Dataset Statistics and Analysis · B SciFormaBench-2K Details · B.1 Structural Inventory

SciFormaData-700K combines broad topic coverage with complex, densely annotated diagrams and editing examples. SciFormaBench-2K assesses structural fidelity using a Components–Arrows–Text inventory and an error taxonomy.

A.2 Dataset Statistics and Analysis

  • SciFormaData-700K draws source papers from 17 ArXiv Computer Science subcategories, covering major fields where methodology diagrams frequently appear.Source: §A.2 · p. 11
  • High- and Medium-complexity samples comprise 38.2% and 48.2% of the dataset, respectively, and the ∼250K High-complexity subset supports Stage-2 SFT.The High-quality fraction reached 45.2% in 2025.Source: §A.2 · p. 11
  • Axis-decomposed captions have a median length of 538 tokens, providing detailed structural descriptions for reconstructing methodology logic from text.Source: §A.2 · p. 11
  • 89.5% of diagrams are landscape with a median aspect ratio of 2.03; 25.9% include formulae, 2.6% are TikZ-rendered, and 71.4% of editing triplets involve 2–3 concurrent modifications.Source: §A.2 · p. 12

B SciFormaBench-2K Details

  • SciFormaBench-2K details cover inventory extraction and an error taxonomy for scoring, using Components, Arrows, and Text as verifiable primitives instead of subjective aesthetic metrics.Source: §B.1 · p. 12, B SciFormaBench-2K Details · p. 12

B.1 Structural Inventory

  • The structural inventory is extracted through joint prompt-image analysis using GPT-5.4 or direct LATEX parsing.The paper refers to Appendix I for the extraction process.Source: §B.1 · p. 12
  • During evaluation, a VLM compares ground-truth and generated images against the Components–Arrows–Text checklist to identify structural discrepancies.Source: §B.1 · p. 12

B.2 Evaluation Taxonomy … C.2 Adaptive Focus Property

The appendix defines severity-weighted, axis-specific structural evaluation and explains how M-DPO directs training toward deficient axes, stops updating once all axes are satisfied, and reduces to standard DPO for one axis.

B.2 Evaluation Taxonomy

  • Critical errors receive weight 1.0 for fundamental misrepresentations or missing elements, while moderate errors receive 0.5 for defective but functional elements.This weighting captures nuanced flaws without penalizing the primary topological skeleton.Source: §B.2 · p. 12
  • Evaluation ignores non-topological traits and uses three axis-specific VLM calls to preserve checking focus and avoid context exhaustion.The protocol provides component-level feedback by assigning each discrepancy an error type and severity score, as illustrated by the structural-inventory checklist.Source: §B.2 · p. 12, Figure 12 · p. 12

C M-DPO Theoretical Analysis

C.1 Gradient Derivation

  • Each gradient step decreases flow-matching loss on the winner and increases it on every axis-anchored loser, weighted by that axis’s w_d.Source: §C.1 · p. 13

C.2 Adaptive Focus Property

  • A failing axis receives large weight, whereas a strongly satisfied axis receives weight near zero.The weights form a softmax over axes parameterized by negated logits.Source: §C.2 · p. 13
  • When all axes are satisfied, M-DPO loss and gradients vanish, and the denominator’s +1 anchor prevents updates from being reassigned arbitrarily.This avoids over-optimization after convergence.Source: §C.2 · p. 13
  • Unlike the per-axis mean objective’s fixed 1/D weights, M-DPO concentrates weight on the bottleneck instead of diluting the gradient for a lone failing axis.Source: §C.2 · p. 13
  • For D=1, M-DPO becomes exactly standard DPO, and for D≥2 it generalizes scalar DPO to conjunctive multi-axis optimization.Source: §C.2 · p. 13

C.3 Multi-Way Bradley-Terry Derivation … E.3 Evaluation Robustness across Multiple Rounds

The supplementary sections formalize M-DPO as conjunctive preference learning with axis-specific negatives, detail SciForma’s base architecture and guarded iterative editing, and test the reliability of SciFormaBench-2K evaluation.

C.3 Multi-Way Bradley-Terry Derivation

  • M-DPO rewards the policy only when the winner is preferred over every axis-specific loser simultaneously.The derivation uses a multi-way Bradley-Terry comparison between one winner and D axis-anchored losers.Source: §C.3 · p. 13

D.2 Iterative Refinement Details

  • SciForma-9B generates up to 3 candidate edits per region and accepts them only if whole-image quality remains within the 0.03 tolerance margin.A global monotonic guard rejects edits that reduce holistic quality beyond the margin and restores the previous figure.Source: Figure 13 · p. 14, §D.2 · pp. 14–15

C.4 Connection to InfoNCE

  • M-DPO is InfoNCE with D negatives, each mined to isolate a specific structural deficiency rather than sampled i.i.d. from the data distribution.The negatives provide informative gradient directions for distinguishing diagrams from axis-specific failure modes.Source: §C.4 · pp. 13–14

D Implementation Details; D.1 Base Model Architecture

  • SciForma uses FLUX.2-klein-base-9B [Labs 2025] with its standard flow-matching DiT, sequence concatenation, and 4D RoPE to support generation and localized editing without task-specific heads.The source image tokens use temporal RoPE offset T=10 while target tokens retain T=0.Source: §D.1 · p. 14

D.2 Iterative Refinement Details

  • The critic prioritizes text defects as critical, module structure as major, and arrows as minor, then localizes edits with a 100×100 grid snapped to a 16×16 latent grid.Minor arrow-direction or style variations are tolerated to avoid unnecessary structural disruption.Source: §D.2 · p. 14

E SciFormaBench-2K In-depth Analysis; E.1 Reflexivity Verification

  • 99.94% of checklist elements were recovered from ground-truth diagrams, with every per-axis and per-tier score above 99.85%.This reflexivity test scores reference diagrams against their own human-verified structural inventory.Source: Table 9 · p. 15, §E.1 · p. 15

E.2 Cross-Evaluator Consistency

  • 61.05 for SciForma-9B + Edit exceeds FLUX.2-dev 32B (48.68), Qwen-Image-2512 (49.76), and GPT-Image-1.5 (59.32) under Qwen3-VL-8B-Instruct scoring.System-level rankings across the two evaluators correlate at Pearson r=0.9945 and Spearman ρ=0.9824.Source: Table 10 · p. 15, §E.2 · p. 15

E.3 Evaluation Robustness across Multiple Rounds

  • Independent GPT-5.4 scoring passes differed by less than 0.22 points, supporting stable, deterministic judgments under the evaluation protocol.The full SciFormaBench-2K evaluation was run twice for a representative subset of models.Source: Table 11 · p. 15, §E.3 · p. 15

E.4 Independence of C, A, T Evaluation Axes … F.2 M-DPO Ablations

The evaluation axes capture distinct failure modes, motivating conjunctive rewards; ablations examine editing supervision and M-DPO choices, with editing triplets improving iterative refinement.

E.4 Independence of C, A, T Evaluation Axes

  • The axis-independence analysis covers four diverse models and 7,391 valid samples, spanning commercial APIs, proprietary systems, open-source diffusion, and intermediate checkpoints.The models are GPT-Image-2, Nano Banana Pro, Wan2.7-Image, and the authors’ intermediate checkpoints.Source: §E.4 · p. 15
  • Component–Arrow correlations are r<0.02 and nonsignificant, while the other pairs have r≤0.26 and shared variance below 7%.This stable pattern supports the rubric's ability to isolate distinct structural properties across model capabilities.Source: §E.4 · pp. 15–16
  • Because axis orthogonality can dilute a scalar average reward and leave arrow errors weakly penalized, the findings motivate per-axis conjunctive bottlenecks.The example is a model that places arrows incorrectly but generates correct components, allowing Component and Text rewards to dominate the average.Source: §E.4 · p. 16

F More Ablations

  • The broader ablation section examines SFT Stage 2 editing data and M-DPO choices, including preference-pair curation, contrastive-axis decomposition, and KL penalty β.The stated purpose is to isolate specific modeling choices.Source: F More Ablations · p. 16

F.1 SFT Ablations

  • For the SFT ablation, two checkpoints trained to 60K steps—with or without 70K editing triplets—receive the same iterative editing budget.SciFormaData-700K contains the editing triplets alongside generation pairs; their benefit is intended to appear through downstream iterative editing.Source: §F.1.1 · p. 16

F.1.1 Editing Triplet Supervision (Stage 2).

  • +2.96 average improvement follows SFT editing-triplet supervision after the same editing budget, with Arrow (+2.44) and Text (+2.66) benefiting most.The gains are consistent across all three axes, supporting localized repair capability for inference-time refinement.Source: §F.1.1 · p. 16

F.2 M-DPO Ablations

  • M-DPO preference-pair curation compares four strategies that vary winner source and rollout diversity on the same SFT checkpoint.This data-construction ablation precedes examination of the axis-specific loss design.Source: §F.2.1 · p. 16

F.2.1 Preference Pair Construction Strategy. … I Prompt Templates

These appendices document preference-data and M-DPO settings, code-based and qualitative comparisons, refinement capabilities and failure modes, and prompts for data construction and evaluation. The prompts assess component, arrow, and text errors and score candidates while filtering layout-only false positives; reported results favor long-short rollouts and triple-axis optimization.

F.2.1 Preference Pair Construction Strategy.

  • +0.83 average: long-short rollouts (4×50+8×25 steps) are the only tested preference-pair strategy to deliver balanced gains across Component, Arrow, and Text.The gains are +0.62 on C, +0.68 on A, and +1.24 on T; ground-truth winners average −1.02, GPT-selected winners −0.48, and uniform short rollouts +0.12. Long and short rollouts expose different structural errors and provide the data for subsequent M-DPO experiments.Source: Table 14 · p. 16, §F.2.1 · p. 16

F.2.2 Number of Contrastive Axes (D)

  • Triple-axis optimization (C, A, T) achieves the highest average, with performance increasing from D=1 to D=2 to D=3.Single-axis optimization can fail to transfer to other dimensions, while dual-axis optimization leaves the third axis lagging; the conjunctive loss concentrates gradients on the weakest axis in each batch.Source: §F.2.1 · pp. 16–17, Table 15 · p. 17

F.2.3 KL Divergence Penalty

  • At β=2000, M-DPO delivers its highest reported gains on the weakest axes: A +1.82 and T +3.16.All three tested β settings clear the SFT baseline across an order of magnitude; Text is most sensitive, with too-small β under-penalizing weak-axis errors and too-large β saturating the gate.Source: Table 16 · p. 17, §F.2.3 · p. 17

G Comparison with Code-Based Baselines

H.1 Generation Gallery

  • SciForma-9B preserves fine-grained spatial relationships better than the three proprietary systems, while open-source baselines show structural, text-rendering, or aesthetic weaknesses.GPT-Image-1.5 uses three fixed canvas sizes, often leaving blank margins; Wan2.7-Image and Nano Banana Pro tend toward heavy module outlines. Qwen-Image-2512 and Z-Image often misrender shapes or arrows, FLUX models poorly render text and spatial structure, and SenseNova-U1 is the strongest open-source baseline but tends toward flat coloring.Source: §H.1.1 · p. 17, §H.1.2 · pp. 17–18

H.4 User Study Details

  • The pairwise preference study samples 50 difficulty-stratified prompts and compares SciForma-9B with Z-Image and GPT-Image-1.5.The sample comprises 17 Simple, 17 Medium, and 16 Hard prompts; annotators compare anonymized generated images alongside the reference on four perceptual axes.Source: §H.4 · p. 18

H.2 Iterative Refinement and Alignment

  • Figures 18 and 19 illustrate localized iterative edits and M-DPO improvements to structural errors, arrow connections, and text layouts.Persistent failures include garbled or horizontalized rotated text, spurious or misattached arrows in dense topologies, and editing artifacts from imprecise bounding-box cropping; proposed directions include rotated-text data, graph-level topology constraints, and context-aware region selection.Source: §H.2 · p. 18, §H.3 · p. 18, Figure 18 · p. 23

I.1 Data Pipeline Prompts

  • The data-construction prompts cover LaTeX macro resolution, diagram filtering, topology-aware captioning, and SAM3 mask verification.The filtering prompts assess target diagram type and quality, while captioning describes global layout, modules, and connections for redrawability.Source: I Prompt Templates · p. 26, §I.1 · pp. 26–27

I.2 Evaluation Prompts

  • Component review checks reference items for absence or incorrectness and prioritizes missing, hallucinated, and distorted elements as critical errors.Structural mismatches, duplicates, and wrong components are moderate errors.Source: §I.2 · p. 27
  • Arrow review flags missing or spurious connections, endpoint-position errors, and wrong components or direction, while excluding routing, curvature, thickness, and color differences.Source: §I.2 · p. 27
  • Text review checks each reference label for presence, legibility, and correctness, treating any character corruption as garbled and not penalizing added text.Missing, garbled, and truncated labels are critical; wrong text and duplicates are moderate errors.Source: §I.2 · p. 28
  • The compact VLM reward scorer compares diagrams for semantic text, arrow, and component errors, disregarding layout, styling, and position changes that preserve logical flow.It requests structured JSON containing diagram descriptions, error categories, and a quality score.Source: §I.2 · p. 28
  • A refinement prompt removes layout-only false positives and retains only true text, connection, and component errors.Source: §I.2 · p. 28

Related papers

Loading 2607.18091v1…