Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A 9-billion-parameter open model can now generate scientific diagrams whose components, arrows, and text are all structurally faithful, beating a strong proprietary image model on the paper's benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 16:01 UTC pith:457R2ZEA

load-bearing objection Serious, well-ablated systems paper; the abstract's no-edit win over GPT-Image-1.5 on the home benchmark flips under their own Qwen rescoring, so the headline needs qualification rather than rejection. the 4 major comments →

arxiv 2607.18091 v1 pith:457R2ZEA submitted 2026-07-20 cs.CV cs.GRcs.LG

SciForma: Structure-Faithful Generation of Scientific Diagrams

Yuxuan Luo , Peng Zhang , Xinjie Zhang , Xun Guo , Zhouhui Lian , Yan Lu This is my paper
classification cs.CV cs.GRcs.LG
keywords scientific diagram generationstructural fidelitycomponent-arrow-text decompositionconjunctive preference optimizationmulti-dimensional DPOdiffusion modelsstructural inventoryverification-gated editing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SciForma aims to make scientific methodology diagrams structurally faithful: every module, arrow, and text label must be correct, because one reversed arrow or unreadable equation invalidates the whole figure. The paper argues that this kind of correctness is conjunctive, so quality should be decomposed into three independently verifiable axes—components, arrows, and text—rather than collapsed into a single scalar score. Its central contribution is M-DPO, a preference-optimization objective that vanishes only when a generated candidate beats an axis-specific loser on all three axes at once, and whose gradient automatically concentrates on whichever axis is failing most. The paper shows that a 9-billion-parameter open model trained this way surpasses all open-source baselines and a strong proprietary image model on both the paper's own structural-fidelity benchmark and an independent reading-comprehension benchmark. If that result holds, open research tools can produce diagrams whose logical structure can be trusted, closing most of the gap with closed commercial systems.

Core claim

Structural fidelity in methodology diagrams is conjunctive: components, arrows, and text must all be right or the figure fails. The paper encodes this in a structural inventory, a checklist extracted from a reference image, used as both training reward and evaluation metric. Its M-DPO objective pairs a winning candidate with one loser per axis and optimizes a multi-way comparison that vanishes only when the winner beats every axis-specific loser, routing gradient weight to the weakest axis. The same inventory gates inference-time edits. The resulting 9B model exceeds all open-source baselines and the proprietary GPT-Image-1.5 on the paper's benchmark, and edges out human-drawn originals on a

What carries the argument

The load-bearing mechanism is M-DPO, a conjunctive preference objective over three axis-anchored preference pairs. For each prompt the system rolls out candidates and labels them under a structural inventory; the overall best candidate is the shared winner and one loser is chosen per axis for being weak on that axis while competitive elsewhere. The loss, derived from a multi-way Bradley–Terry model, is log(1 + sum_d exp(-beta * Delta_d)), where Delta_d is the relative preference margin for axis d. It vanishes only if the winner beats every axis-specific loser, and its per-axis gradient weight is a softmax over deficits, so the strongest update goes to the axis with the largest failure. This

Load-bearing premise

The whole pipeline assumes that a vision-language model's checklist of components, arrows, and text is a complete and unbiased measure of whether a diagram is structurally correct.

What would settle it

Rescore the benchmark outputs using human annotators who verify every component, arrow, and label against the reference, adding dimensions the inventory omits (rotated text, color semantics, layout aesthetics). If human per-axis error rates fail to reproduce the ordering SciForma-9B over the proprietary model, or if an adversarially constructed diagram that passes all component/arrow/text checklists is nevertheless judged structurally invalid by a reader, then the inventory-based verification is incomplete and the claimed structural-fidelity gain is an artifact of the evaluator.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A 9-billion-parameter open model can match or beat much larger and closed systems on structurally faithful diagram generation, so publication-quality diagram generation no longer requires proprietary APIs.
  • Because the reward and the evaluation share the same explicit checklists, failures become diagnosable: a low Arrow score tells the user exactly which arrows are missing or misrouted, enabling targeted repair.
  • Scalar post-training methods—averaging rewards or running online RL on one number—plateau or regress on these diagrams; the conjunctive formulation is what breaks the plateau, with the largest gains on the weakest axes.
  • Inference-time editing with a verification gate and rollback monotonically improves the output, so generation and correction can be unified in one model.
  • Pixel-space generation with structural supervision outperforms code-based diagram synthesis by roughly a factor of two on the home benchmark.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • As an editorial extension, the same 'conjunctive axes plus weakest-axis gradient routing' recipe could transfer to any domain where a single error invalidates a visual artifact—circuit schematics, UML models, game levels—provided a checklist can be extracted.
  • Because M-DPO's reward signal and the home benchmark's ranking both come from vision-language checklists, part of the reported gain may reflect alignment with that particular evaluator's notion of structure; an independent human-annotated error study would clarify how much is genuine topological understanding.
  • A testable extension is to apply the same conjunctive objective to long-form structured outputs such as procedures or instructions, where scalar rewards are known to dilute rare but critical errors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. SciForma introduces a structural-inventory formalism for scientific methodology diagrams, decomposing diagram quality into Component (C), Arrow (A), and Text (T) axes. The paper contributes (1) SciFormaData-700K, a large curated generation/editing dataset; (2) SciFormaBench-2K, a 2,000-sample benchmark scored by VLM checklist verification; (3) M-DPO, a multi-dimensional conjunctive preference optimization objective that couples axis-specific losers through a multi-way Bradley–Terry loss; and (4) an inference-time iterative refinement loop with verification-gated editing. The main empirical claim is that SciForma-9B, a 9B open model fine-tuned from FLUX.2-klein-base-9B, exceeds all open-source baselines and GPT-Image-1.5 on both SciFormaBench-2K and AIBench, with additional support from a 36-subject user study. The paper also reports ablations showing M-DPO improves over SFT and over scalar or mean-DPO alternatives.

Significance. If the claims hold, this is a substantive contribution: a large, professionally relevant dataset; a benchmark with element-level, axis-decomposed structural verification; and a training objective that explicitly models conjunctive correctness. The paper is unusually thorough in its supporting evidence: an independent AIBench evaluation, a Qwen3-VL rescoring of the full leaderboard, a reflexivity check on ground-truth diagrams, multi-round judge-stability measurements, and an analysis of C/A/T axis orthogonality. The M-DPO formulation is clean and reduces to standard DPO when D=1. These strengths make the paper a candidate for publication, but the headline claim is currently overstated relative to the paper’s own cross-evaluator results.

major comments (4)
  1. [§5.2 vs. Appendix E.2, Tables 2a and 10] The central claim that no-edit SciForma-9B exceeds GPT-Image-1.5 on SciFormaBench-2K is evaluator-dependent. With GPT-5.4, SciForma-9B scores 69.51 vs. GPT-Image-1.5's 68.96 (+0.55). Under the paper's own Qwen3-VL rescoring (Table 10), the same pair reverses: SciForma-9B 59.21 vs. GPT-Image-1.5 59.32 (−0.11). Only SciForma-9B+Edit (61.05) overtakes GPT-Image-1.5 under the independent evaluator. The statement in E.2 that 'the significance of our gains remains completely unaffected' is not supported: a rank flip at the headline comparison is not captured by the aggregate correlation r=0.9945. The AIBench margin is independent and survives, but the SciFormaBench-2K claim in the abstract and Section 5.2 should be qualified, or a direct head-to-head with a non-overlapping judge should be provided.
  2. [§3.2, §4.5.1, §5.2] There is a circularity risk: the M-DPO reward signal (Qwen3-VL scoring on the C/A/T inventory) and the SciFormaBench-2K evaluation (GPT-5.4 scoring on the same inventory) are both built from the same structural checklist. The reflexivity check (Table 9) shows the evaluator can find elements it is told to look for, but it does not establish that the checklist is a complete or unbiased proxy for structural fidelity. The paper should explicitly test whether an independent, non-VLM or human structural check (e.g., a human judge re-extracting C/A/T from generated images, or a layout/color-semantics audit) corroborates the benchmark ordering, especially for the SciForma-9B vs. GPT-Image-1.5 comparison.
  3. [§5.2, Tables 2a/2b] No confidence intervals or significance tests are reported for the main benchmark comparisons. The headline margins are small: +0.55 on SciFormaBench-2K (GPT-5.4) and +0.20 over human originals on AIBench. Given per-sample variance and stochastic VLM scoring, bootstrap confidence intervals or paired tests over the 2,000/300 samples should be reported. Without them, the reader cannot distinguish a real improvement from evaluator noise.
  4. [§5.7 vs. Appendix H.4] The user study description is inconsistent. The main text reports 36 subjects and 30 samples, while Appendix H.4 describes 50 prompts and majority voting by three annotators per pair. The appendix also says Fleiss' kappa is reported, but no kappa value appears. Since the user study is one of the three supporting lines for the benchmark, this discrepancy needs to be reconciled and the inter-annotator agreement actually reported.
minor comments (6)
  1. [Abstract/Section 5.2] The phrase 'SciForma-9B exceeds ... GPT-Image-1.5' should be clarified to distinguish the no-edit model from SciForma-9B+Edit, since the no-edit model's lead reverses under the independent evaluator.
  2. [Appendix E.2] The claim of 'near-perfect correlation' (r=0.9945) is not an adequate substitute for reporting the specific SciForma-9B vs. GPT-Image-1.5 comparison under both evaluators. Consider adding a paired-difference table.
  3. [§5.5, Table 4] The GDRO baseline is described as ranking candidates with a single scalar reward; the abbreviation is potentially confusing with standard group distributionally robust optimization. A one-sentence clarification of the baseline's objective would help.
  4. [Appendix I.2] The reward-scoring prompt and the benchmark evaluation prompts are similar in spirit but not identical; the paper should state explicitly that the same prompt template was not used for reward construction and evaluation, or discuss any effect of prompt overlap.
  5. [§4.6 and Appendix D.2] The monotonic guard is defined as S_new >= S_old - epsilon with epsilon=0.03, but S is not precisely defined. The paper should state which score (average, per-axis, or a composite) is used for the guard.
  6. [§5.2, Table 2a] GPT-Image-2 is listed as a proprietary baseline, but its margin over SciForma-9B+Edit is large (85.62 vs. 72.40). The paper's contribution would be clearer if the discussion acknowledged this residual gap rather than only emphasizing the GPT-Image-1.5 comparison.

Circularity Check

0 steps flagged

No significant circularity: the derivation chain is self-contained and independently grounded.

full rationale

The paper's load-bearing derivations are (i) the structural-inventory formalism (Sec. 3), (ii) the M-DPO objective Eq. (4) with its gradient reweighting (Sec. 4.5.2, App. C), and (iii) the empirical comparisons (Sec. 5). None of these reduces to its own inputs by construction. Eq. (4) is derived from a multi-way Bradley-Terry model and reduces to standard DPO when D=1; the adaptive weights in Eq. (5) follow by differentiation. The C/A/T inventory is an external operationalization anchored to UML/BPMN standards, not defined by the model's outputs. The shared rubric between the Qwen3-VL-based M-DPO reward and the GPT-5.4-scored SciFormaBench-2K is a metric-alignment caveat, not a circular derivation: the reward and the leaderboard use different judges, and the paper provides independent grounding via Qwen3-VL rescoring (Table 10), AIBench (Sec. 5.3), PaperBanana (Sec. 5.4), and a user study with r=0.76 (Sec. 5.7). The GPT-5.4/Qwen3-VL rank flip on the specific SciForma-9B vs GPT-Image-1.5 pair is an evaluator-robustness concern, which belongs under correctness risk, not circularity. No load-bearing self-citation chain or imported uniqueness theorem is present, and the paper openly lists its own limitations (proprietary VLM dependence, resolution, auto-generated inventories).

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

M-DPO's math is self-contained, but the empirical contribution rests on several borrowed constructs: the UML/BPMN-inspired three-axis decomposition, VLM extraction of checklists, and the choice of reference diagrams as ground truth. The main free parameters are the preference-optimization hyperparameters and the hand-set severity weights that define the benchmark score itself. The only invented "entity" in the paper is the structural inventory abstraction, which has an external handle through human verification and user studies.

free parameters (7)
  • M-DPO temperature β = 2000
    KL/DPO scale in Eq. (4); swept in Table 16, 2000 used for all main results. Controls how strongly per-axis margins are amplified.
  • Preference margin thresholds δ_min, δ_max = 0.25, 0.60
    Section 5.1; filter rollouts into valid (1+D) tuples; determine which 16K of 50K prompts enter M-DPO training.
  • Winner gate τ = 0.70
    Section 5.1; minimum score for selecting shared winner; changes preference pair distribution.
  • Error severity weights w_e = critical 1.0, moderate 0.5
    Section 3.2 Eq. (1) and Table 8; hand-set rubric weights define every SciFormaBench score, including headline comparisons.
  • Iterative-refinement tolerance ε = 0.03
    Appendix D.2; global guard accepts edit only if S_new ≥ S_old − ε; controls edit-loop gains (+2.89).
  • Inference hyperparameters = CFG=4.0, 50 Euler steps
    Section 5.1; standard but chosen; affects all generated outputs and benchmark scores.
  • Rollout budget K = 12 (4×50-step, 8×25-step)
    Section 5.1 and Table 14; ablation shows long-short mixing is required for gains; cost/benefit choice.
axioms (5)
  • domain assumption Diagram quality is fully captured by C/A/T primitives (Component, Arrow, Text) drawn from UML/BPMN.
    Section 3.1 adopts UML/BPMN grammar and asserts these three axes are the structural inventory; if other dimensions are load-bearing, the reward and benchmark miss them.
  • domain assumption A VLM-generated structural inventory from a reference image is a faithful, complete ground-truth checklist.
    Section 3.2 and Appendix E.1 rely on VLM extraction; reflexivity (99.94%) supports but does not prove unbiasedness for generated images.
  • domain assumption The reference diagram is the correct ground truth; outputs differing from it are wrong even if prompt-compliant.
    Benchmark Eq. (1) scores outputs against the reference checklist; a different but equally valid diagram from the same prompt is penalized. Section B.1 defines inventory via prompt-image analysis of the reference.
  • standard math Multi-way Bradley-Terry preference model governs conjunctive choices across axes.
    Section 4.5.2 and Appendix C.3 derive M-DPO from a (1+D)-way BT probability; this is a modeling choice, not a theorem about actual human preferences.
  • domain assumption Implicit-reward margins Δ_d from Qwen3-VL scores are a valid optimization target for structural quality.
    M-DPO assumes the VLM score margins are a valid reward signal; if Qwen3-VL scores are gameable, the policy can improve the score without improving structure.
invented entities (1)
  • Structural inventory (C/A/T checklist) independent evidence
    purpose: Ground-truth checklists for per-axis training rewards, benchmark scoring, and inference-time defect detection.
    Operationalized via prompt templates; human reflexivity (99.94%) and user-study correlation (r=0.76) provide an external handle, but the construct itself is defined by the authors.

pith-pipeline@v1.3.0-alltime-deepseek · 32783 in / 16048 out tokens · 144463 ms · 2026-08-01T16:01:03.100800+00:00 · methodology

0 comments
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.18091 by Peng Zhang, Xinjie Zhang, Xun Guo, Yan Lu, Yuxuan Luo, Zhouhui Lian.

Figure 1
Figure 1. Figure 1: Diverse scientific methodology diagrams generated by SciForma. Our framework ensures strict [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SciForma. (a) Pipeline: Two-stage SFT yields SciForma-Base, followed by M-DPO to produce SciForma-9B. (b) Training Details: Left: SFT data formats for joint generation and editing. Right: M-DPO optimizes a shared winner against dimension-anchored losers to model conjunctive preference. (c) Iterative Refinement: A closed-loop inference pipeline detects structural defects and applies localized ed… view at source ↗
Figure 1
Figure 1. Figure 1: Fig1 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the SciFormaData-700K construction. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: M-DPO Algorithm Overview. M-DPO computes axis-specific mar [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training trajectories from the same 90K SFT checkpoint on SciFormaBench-2K (GPT-5.4). Red: M-DPO; Blue: continued SFT. M-DPO achieves larger [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: User study. Win/Tie/Lose percentages for SciForma-9B against Nano Banana Pro and Wan2.7-Image across four evaluation dimensions. distortion. The resulting diagrams show cleaner, more coherent topologies that closely align with the ground truth. 5.7 User Study We further conduct a user study, comparing SciForma-9B with Wan2.7-Image and Nano Banana Pro. Thirty-six graduate students specialized in AI evaluate… view at source ↗
Figure 7
Figure 7. Figure 7: SciForma-9B generates high-fidelity pipeline figures with superior structural integrity and precise text rendering. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the closed-loop iterative refinement process. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of SciForma-9B on PaperBananaBench. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual improvements introduced by Multi-Dimensional Conjunctive Preference Optimization. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: SciFormaData-700K statistics. From left to right: temporal growth of [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustration of structural inventory verification on a generated dia [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Anatomy of the Closed-Loop Editor. Stage 1: The GPT-5.4 critic uses a 100×100 visual grid and priority hierarchy to localize defects. The raw bounding box is then refined and snapped to the 16×16 VAE layout. Stage 2: SciForma-9B dynamically generates candidate edits, which are subjected to a global monotonic guard (Snew ≥ Sold − 𝜖). Updates that disrupt holistic layout quality beyond the tolerance margin … view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison with proprietary models (1/2). SciForma-9B vs. Wan2.7-Image, GPT-Image-1.5, and Nano Banana Pro on the same prompts. [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison with proprietary models (2/2). Same setup as Figure 14. [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison with open-source models (1/2). SciForma-9B vs. SenseNova-U1, Qwen-Image-2512, and Z-Image. [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparison with open-source models (2/2). SciForma-9B vs. FLUX.2-dev-32B and FLUX.2 Klein Base-9B. [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Further examples of iterative visual refinement. SciForma-9B accurately updates localized structures, texts, and arrow connections based on editing [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Qualitative effects of M-DPO. The preference-aligned model (SciForma-9B + M-DPO) effectively corrects structural errors, repairs arrow connections, [PITH_FULL_IMAGE:figures/full_fig_p024_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Representative failure cases of SciForma-9B. (a) Rotated text: verti [PITH_FULL_IMAGE:figures/full_fig_p025_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: User study questionnaire screenshot. (a) Study instructions out [PITH_FULL_IMAGE:figures/full_fig_p025_21.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing

    cs.CV 2026-07 conditional novelty 5.0

    A compact 4B image generation/editing system with a fast one-step VAE, native-resolution packing, RL alignment, and 4-step distillation reports competitive benchmarks against 6B–80B open models.

Reference graph

Works this paper leans on

121 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    2025 , url=

    Black Forest Labs , title=. 2025 , url=

    2025

  2. [2]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  3. [3]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  4. [4]

    2025 , eprint=

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space , author=. 2025 , eprint=

    2025

  5. [5]

    2306.00800 , archivePrefix=

    Figgen: Text to scientific figure generation , author=. 2306.00800 , archivePrefix=

    Pith/arXiv arXiv

  6. [6]

    2505.02567 , archivePrefix=

    Unified multimodal understanding and generation models: Advances, challenges, and opportunities , author=. 2505.02567 , archivePrefix=

    arXiv

  7. [7]

    2506.23044 , archivePrefix=

    Ovis-u1 technical report , author=. 2506.23044 , archivePrefix=

    Pith/arXiv arXiv

  8. [8]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Luo, Yuxuan and Tang, Jiaqi and Huang, Chenyi and Hao, Feiyang and Lian, Zhouhui , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  9. [9]

    2511.16719 , archivePrefix=

    Sam 3: Segment anything with concepts , author=. 2511.16719 , archivePrefix=

    Pith/arXiv arXiv

  10. [10]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    SciDoc2Diagrammer-MAF: Towards generation of scientific diagrams from documents guided by multi-aspect feedback refinement , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=. 2024 , publisher=

    2024

  11. [11]

    The Fourteenth International Conference on Learning Representations , year=

    TikZilla: Scaling Text-to-TikZ with High-Quality Data and Reinforcement Learning , author=. The Fourteenth International Conference on Learning Representations , year=

  12. [12]

    AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ , url =

    Belouadi, Jonas and Lauscher, Anne and Eger, Steffen , booktitle =. AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ , url =. 2024 , publisher =

    2024

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Detikzify: Synthesizing graphics programs for scientific figures and sketches with tikz , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Belouadi, Jonas and Ilg, Eddy and Keuper, Margret and Tanaka, Hideki and Utiyama, Masao and Dabre, Raj and Eger, Steffen and Ponzetto, Simone , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  15. [15]

    2310.12128 , archivePrefix=

    Diagrammergpt: Generating open-domain, open-platform diagrams via llm planning , author=. 2310.12128 , archivePrefix=

    Pith/arXiv arXiv

  16. [16]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    From words to structured visuals: A benchmark and framework for text-to-diagram generation and editing , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=. 2025 , publisher=

    2025

  17. [17]

    2602.03828 , archivePrefix=

    AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations , author=. 2602.03828 , archivePrefix=

    arXiv

  18. [18]

    2603.06674 , archivePrefix=

    AutoFigure-Edit: Generating Editable Scientific Illustration , author=. 2603.06674 , archivePrefix=

    arXiv

  19. [19]

    2601.23265 , archivePrefix=

    PaperBanana: Automating Academic Illustration for AI Scientists , author=. 2601.23265 , archivePrefix=

    arXiv

  20. [20]

    2601.04390 , archivePrefix=

    SciFig: Towards Automating Scientific Figure Generation , author=. 2601.04390 , archivePrefix=

    Pith/arXiv arXiv

  21. [21]

    DiagramEval: Evaluating

    Liang, Chumeng and You, Jiaxuan , booktitle=. DiagramEval: Evaluating. 2025 , address =. doi:10.18653/v1/2025.emnlp-main.640 , pages =

    work page doi:10.18653/v1/2025.emnlp-main.640 2025

  22. [22]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages=

    SCISKETCH: An Open-source Framework for Automated Schematic Diagram Generation in Scientific Papers , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages=. 2025 , publisher=

    2025

  23. [23]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Svgen: Interpretable vector graphics generation with large language models , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=. 2025 , publisher=

    2025

  24. [24]

    2511.18036 , archivePrefix=

    Paper2SysArch: Structure-Constrained System Architecture Generation from Scientific Papers , author=. 2511.18036 , archivePrefix=

    arXiv

  25. [25]

    2603.28068 , archivePrefix=

    AIBench: Evaluating Visual-Logical Consistency in Academic Illustration Generation , author=. 2603.28068 , archivePrefix=

    arXiv

  26. [26]

    2604.04172 , archivePrefix=

    GENFIG1: Visual Summaries of Scholarly Work as a Challenge for Vision-Language Models , author=. 2604.04172 , archivePrefix=

    Pith/arXiv arXiv

  27. [27]

    2604.08211 , archivePrefix=

    SciFigDetect: A Benchmark for AI-Generated Scientific Figure Detection , author=. 2604.08211 , archivePrefix=

    Pith/arXiv arXiv

  28. [28]

    2512.12220 , archivePrefix=

    ProImage-Bench: Rubric-Based Evaluation for Professional Image Generation , author=. 2512.12220 , archivePrefix=

    arXiv

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Mmmg: A massive, multidisciplinary, multi-tier generation benchmark for text-to-image reasoning , author=. Advances in Neural Information Processing Systems , volume=. 2025 , publisher=

    2025

  30. [30]

    2602.09809 , archivePrefix=

    SciFlow-Bench: Evaluating Structure-Aware Scientific Diagram Generation via Inverse Parsing , author=. 2602.09809 , archivePrefix=

    Pith/arXiv arXiv

  31. [31]

    2507.02212 , archivePrefix=

    Sciga: A comprehensive dataset for designing graphical abstracts in academic papers , author=. 2507.02212 , archivePrefix=

    Pith/arXiv arXiv

  32. [32]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Multimodal arxiv: A dataset for improving scientific comprehension of large vision-language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2024 , publisher=

    2024

  33. [33]

    2604.14941 , archivePrefix=

    Text2Arch: A Dataset for Generating Scientific Architecture Diagrams from Natural Language Descriptions , author=. 2604.14941 , archivePrefix=

    Pith/arXiv arXiv

  34. [34]

    2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=

    Identifying the central figure of a scientific paper , author=. 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=. 2019 , publisher=

    2019

  35. [35]

    2407.04903 , archivePrefix=

    Mmsci: A dataset for graduate-level multi-discipline multimodal scientific understanding , author=. 2407.04903 , archivePrefix=

    Pith/arXiv arXiv

  36. [36]

    Workshop on Responsibly Building the Next Generation of Multimodal Foundational Models , year=

    Bigdocs: An open and permissively-licensed dataset for training multimodal models on document and code tasks , author=. Workshop on Responsibly Building the Next Generation of Multimodal Foundational Models , year=

  37. [37]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  38. [38]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Diffusion model alignment using direct preference optimization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2024 , publisher=

    2024

  39. [39]

    2402.03300 , archivePrefix=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. 2402.03300 , archivePrefix=

    Pith/arXiv arXiv

  40. [40]

    Training Diffusion Models with Reinforcement Learning , url =

    Black, Kevin and Janner, Michael and Du, Yilun and Kostrikov, Ilya and Levine, Sergey , booktitle =. Training Diffusion Models with Reinforcement Learning , url =. 2024 , publisher =

    2024

  41. [41]

    2601.02036 , archivePrefix=

    GDRO: Group-level Reward Post-training Suitable for Diffusion Models , author=. 2601.02036 , archivePrefix=

    arXiv

  42. [42]

    Sensors , volume=

    Flow-Multi: A Flow-Matching Multi-Reward Framework for Text-to-Image Generation , author=. Sensors , volume=

  43. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Margin-Aware Preference Optimization for Aligning Diffusion Models without Reference , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2026 , publisher=

    2026

  44. [44]

    , author=

    Gemini 3 Pro Image (Nano Banana Pro): Define the next generation of professional asset creation. , author=. 2025 , howpublished=

    2025

  45. [45]

    2026 , howpublished=

    2026

  46. [46]

    2025 , note =

    Introducing 4o Image Generation , howpublished =. 2025 , note =

    2025

  47. [47]

    2026 , note =

    Introducing ChatGPT Images 2.0 , howpublished =. 2026 , note =

    2026

  48. [48]

    2025 , note =

    Deeper Thinking, More Accurate Generation: Introducing. 2025 , note =

    2025

  49. [49]

    2024 , note =

    Alibaba Cloud , title =. 2024 , note =

    2024

  50. [50]

    2024 , howpublished=

    Black Forest Labs , title=. 2024 , howpublished=

    2024

  51. [51]

    2025 , eprint=

    Qwen-Image Technical Report , author=. 2025 , eprint=

    2025

  52. [52]

    2511.22699 , archivePrefix=

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer , author=. 2511.22699 , archivePrefix=

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:2512.07584 , year=

    Longcat-image technical report , author=. arXiv preprint arXiv:2512.07584 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2505.14683 , year =

    Emerging Properties in Unified Multimodal Pretraining , author =. arXiv preprint arXiv:2505.14683 , year =

    Pith/arXiv arXiv

  55. [55]

    International Conference on Machine Learning , year=

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author=. International Conference on Machine Learning , year=

  56. [56]

    2210.02747 , archivePrefix=

    Flow Matching for Generative Modeling , author=. 2210.02747 , archivePrefix=

    Pith/arXiv arXiv

  57. [57]

    European Conference on Computer Vision , pages=

    PixArt- : Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation , author=. European Conference on Computer Vision , pages=

  58. [58]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2024 , publisher=

    2024

  59. [59]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  60. [60]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Scalable Ranked Preference Optimization for Text-to-Image Generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2025 , publisher=

    2025

  61. [61]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  62. [62]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Learning Multidimensional Human Preference for Text-to-Image Generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2024 , publisher=

    2024

  63. [63]

    Advances in

    Liu, Zexiang and He, Xianglong and Li, Yangguang , eprint=. Advances in

  64. [64]

    2604.13029 , archivePrefix=

    Visual Preference Optimization with Rubric Rewards , author=. 2604.13029 , archivePrefix=

    Pith/arXiv arXiv

  65. [65]

    2511.20651 , archivePrefix=

    RubricRL: Simple Generalizable Rewards for Text-to-Image Generation , author=. 2511.20651 , archivePrefix=

    Pith/arXiv arXiv

  66. [66]

    2604.11626 , archivePrefix=

    RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time , author=. 2604.11626 , archivePrefix=

    Pith/arXiv arXiv

  67. [67]

    Zhang, Liyu and Li, Kehan and Han, Tingrui and Zhao, Tao and Sheng, Yuxuan and He, Shibo and Li, Chao , eprint=

  68. [68]

    SuperFlow: Training Flow Matching Models with

    Chen, Kaijie and Xu, Zhiyang and Shen, Ying and Lin, Zihao and Yao, Yuguang and Huang, Lifu , eprint=. SuperFlow: Training Flow Matching Models with

  69. [69]

    ACM Computing Surveys , year=

    Alignment of Diffusion Models: Fundamentals, Challenges, and Future , author=. ACM Computing Surveys , year=

  70. [70]

    Advances in Neural Information Processing Systems , volume=

    ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation , author=. Advances in Neural Information Processing Systems , volume=. 2023 , publisher=

    2023

  71. [71]

    2306.09341 , archivePrefix=

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis , author=. 2306.09341 , archivePrefix=

    Pith/arXiv arXiv

  72. [72]

    Advances in Neural Information Processing Systems , volume=

    Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation , author=. Advances in Neural Information Processing Systems , volume=. 2023 , publisher=

    2023

  73. [73]

    2025 , eprint=

    Qwen3-VL Technical Report , author=. 2025 , eprint=

    2025

  74. [74]

    2025 , note=

    Introducing. 2025 , note=

    2025

  75. [75]

    2026 , note=

    Introducing. 2026 , note=

    2026

  76. [76]

    2409.18839 , archivePrefix=

    Mineru: An open-source solution for precise document content extraction , author=. 2409.18839 , archivePrefix=

    Pith/arXiv arXiv

  77. [77]

    2603.18528 , archivePrefix=

    Correlation-Weighted Multi-Reward Optimization for Compositional Generation , author=. 2603.18528 , archivePrefix=

    Pith/arXiv arXiv

  78. [78]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Aligning Text-to-Image Diffusion Models With Constrained Reinforcement Learning , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  79. [79]

    The Fourteenth International Conference on Learning Representations , year=

    Sample Reward Soups: Query-efficient Multi-Reward Guidance for Text-to-Image Diffusion Models , author=. The Fourteenth International Conference on Learning Representations , year=

  80. [80]

    2601.03468 , archivePrefix=

    Understanding Reward Hacking in Text-to-Image Reinforcement Learning , author=. 2601.03468 , archivePrefix=

    arXiv

Showing first 80 references.