Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Text-to-image models systematically violate physical commonsense, and a knowledge-graph-driven benchmark with dual-path verification diagnoses these failures, while batch-aggregated prompt optimization measurably corrects some of them.

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 01:49 UTC pith:XUBIZMTJ

load-bearing objection The OmniPhys benchmark is a real, citable resource; the OmniPrompt gains are plausible but not yet shown to be physics rather than VLM preference. the 4 major comments →

arxiv 2607.25641 v1 pith:XUBIZMTJ submitted 2026-07-28 cs.CV cs.AI

OmniPhys: Knowledge-Graph-Driven Benchmarking and Collective Optimization for Physical Commonsense in Text-to-Image Generation

Yajing Xu , Yarong Lan , Jiaoyan Chen , Yichi Zhang , Jeff Z. Pan , Mingchen Tu , Zhizhen Liu , Wen Zhang
show 1 more author
Huajun Chen
This is my paper
classification cs.CV cs.AI
keywords physical commonsensetext-to-image generationknowledge graphbenchmarkprompt optimizationvision-language model evaluationgradient hallucinationcollective feedback
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.

The paper argues that text-to-image models fail basic physics not because of a single weak spot but because they lack internalized physical principles, and that existing benchmarks are too coarse to say which principles are missing. To fix this, it builds OmniPhys, a benchmark of 1,551 implicit prompts anchored to 14 physical knowledge points (e.g., buoyancy, refraction, communicating vessels) extracted from a knowledge graph aligned with PhET simulations and standard curricula. Each sample is verified with a dual-path protocol that requires both a discriminative VQA answer and a descriptive consistency statement, so a lucky guess does not count. It then proposes OmniPrompt, which improves prompts by aggregating feedback over 4 images per query and batches of 2–12 queries per update, turning physical alignment into a discrete optimization over natural-language meta-policies. The reported result is that this collective optimization improves physical consistency on three backbones (FLUX.1-dev, JanusPro-7B, Lumina-Image-2.0) over the instance-level TextGrad baseline, and that the evolved policies transfer to closed-source engines.

Core claim

The paper's central claim is that physical commonsense in text-to-image generation can be systematically diagnosed and improved. OmniPhys establishes that 12 representative models—including frontier closed-source engines—share the same bottlenecks: high VQA scores but low DCS scores, meaning models can answer simple yes/no probes but fail stricter descriptive consistency checks; optical phenomena like reflection and structural constraints like the leverage principle are nearly universally wrong. The second claim is that OmniPrompt's batch-aggregated linguistic gradients resolve 'gradient hallucinations,' and that the evolved meta-policies transfer across backbones, e.g., raising joint scores

What carries the argument

The Physical Knowledge Graph (PKG): a 33-node, 32-edge hierarchy with 3 domains, 7 concepts, 9 principles, and 14 leaf physical knowledge points (PKPs), each annotated with atomic physical statements that admit binary visually verifiable outcomes. The dual-path verification protocol: for each implicit prompt, an LLM generates n coupled (VQA, DCS) probes; an image passes only if both the discriminative VQA answers match gold answers and all descriptive consistency statements hold, computed as a joint score averaged over 4 images. The OmniPrompt loop: a prompt enhancer LLM expands each query under a natural-language meta-policy P; the T2I model draws K=4 images; a multimodal evaluator produces

Load-bearing premise

The load-bearing premise is that the automated vision-language evaluator is a trustworthy oracle for physical consistency in static 2D images, which the paper's own human evaluation supports only at 80% agreement with moderate inter-annotator reliability (AC1=0.637); if the evaluator is systematically wrong, the benchmark scores and the reported optimization gains are unreliable.

What would settle it

Re-score a random sample of OmniPrompt- and TextGrad-generated images with human physics-trained annotators instead of the VLM, and check whether the reported Joint Score advantage of OmniPrompt persists; alternatively, audit the evolved meta-policy for outcome-revealing words (e.g., 'floats upright', 'reflected at equal angles') that would inflate VQA/DCS scores by leaking the expected answer into the prompt.

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

If this is right

  • If a model fails a PKP, the pair of VQA and DCS probes pinpoints the specific law (e.g., equal liquid levels in communicating vessels) rather than just a broad 'mechanics' category.
  • Prompt-level optimization can recover some physical consistency without fine-tuning model weights, which is cheaper and works on closed-source engines.
  • Batch aggregation over queries is the key to filtering seed noise; single-query optimization (TextGrad) is unstable and can underperform even a zero-shot CoT-augmented prompt (as on Lumina-Image-2.0).
  • The VQA-versus-DCS gap implies that any evaluation relying on a single binary probe will overestimate physical consistency; the dual-path protocol is necessary for trustworthy scores.
  • Evolved meta-policies transfer across backbones and synthesis paradigms (diffusion, unified multimodal, autoregressive, closed-source), suggesting the learned instructions capture general physical-reasoning patterns.

Where Pith is reading between the lines

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

  • The paper trains its meta-policy on only 37 prompts; whether the optimization finds genuinely general physical heuristics or just overfits to those few scenarios is not directly tested, though the transfer results to closed-source engines are suggestive.
  • Because the same style of VLM that drives training feedback also scores the final test set, the reported gains are partly circular if the VLM has systematic biases (e.g., rewarding prompts that explicitly describe expected outcomes); the paper's human evaluation showing 80% agreement leaves a 20% error margin that is never analyzed for its effect on the scores.
  • The method could plausibly extend beyond physics to any structured commonsense domain (chemistry, biology, tool use) that can be encoded as a knowledge graph with atomic, visually verifiable statements.
  • The binary all-or-nothing scoring punishes partially correct images; a graded score might reveal that OmniPrompt improves partial correctness even when joint scores move little.

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 / 5 minor

Summary. The paper introduces OmniPhys, a benchmark of 1,551 implicit physical-commonsense prompts grounded in a Physical Knowledge Graph with 14 physical knowledge points, together with a dual-path VQA/DCS verification protocol scored by a VLM. It further presents OmniPrompt, an iterative prompt-optimization framework that aggregates feedback over K images per query and B queries per update to evolve a linguistic meta-policy via TextGrad. Experiments across 12 T2I models report universal physical bottlenecks, and OmniPrompt is claimed to improve physical consistency over raw queries, CoT-Aug, and TextGrad baselines on FLUX.1-dev, JanusPro-7B, and Lumina-Image-2.0.

Significance. The benchmark construction is systematic: grounding in PhET simulations and a PKG, separating implicit queries from atomic statements, and human-validating probe pairs (96–100% validity) are strengths. If the optimization result holds, the batch-aggregated linguistic-gradient approach would be a practical, training-free way to improve physical consistency. However, the central efficacy claim currently rests on a VLM judge that is also the training signal; the 80% human–VLM agreement and small effect sizes leave the claim vulnerable to evaluator-bias artifacts. The paper's contribution is therefore valuable but not yet fully established.

major comments (4)
  1. [§5.4 (Table 4) and §4.2.3 / §3.2] The VLM (Gemini-2.5-Pro) serves as both the training-time feedback signal (Eq. 3) and the final scoring oracle (Joint Score in §3.2). The human evaluation in Table 4 reports only 80.0% human–VLM agreement with inter-annotator AC1=0.637 on image labeling, and was conducted on 200 images generated from raw/CoT-Aug prompts, not from OmniPrompt-optimized prompts. Because the meta-policy is explicitly optimized to maximize this same VLM's DCS scores, the small gains in Table 3 (0.010–0.025) could reflect VLM biases (e.g., sensitivity to prompt wording or length) rather than genuine physical improvement. To support the 'significant enhancement' claim, the authors must provide human-scored evaluation on the final optimized prompts, or show that the VLM-based gains align with human judgments on a held-out set of optimized outputs.
  2. [Table 3] No confidence intervals, standard errors, or significance tests are reported for any of the OmniPrompt gains. The differences over TextGrad are 0.012 (FLUX), 0.010 (JanusPro), and 0.025 (Lumina) in Joint Score. Given the stochasticity of T2I sampling (K=4 images per prompt) and the moderate reliability of the VLM scorer (AC1=0.637), these differences are within a range that could arise from sampling noise. Please report variance across independent seeds or bootstrap confidence intervals, and a paired significance test per backbone.
  3. [§5.1.4 / Fig. 9] The final batch size B used for the main OmniPrompt results in Table 3 is not stated. The sensitivity analysis in Figure 9 shows that B substantially affects performance (e.g., FLUX.1-dev has a sharp drop outside [6,10]), so the reported results are uninterpretable without knowing B. State the B used for each reported result and justify the choice; also report sensitivity for all backbones, not only JanusPro-7B.
  4. [§4.2.3 / §3.2] Training feedback uses DCS probes only (§4.2.3: 'E provides DCS-only feedback'), while the final evaluation uses the Joint Score that multiplies VQA and DCS. This mismatch is not discussed. In Table 3, VQA scores are essentially unchanged on FLUX and JanusPro, indicating that the optimization improves only the DCS path. The claim that OmniPrompt enhances 'physical consistency' more broadly should be qualified, and the authors should explain why VQA is excluded from training or discuss the implications for the dual-path claim.
minor comments (5)
  1. [§5.2 / Table 2] Table 2 reports scores on the full 1,551-prompt benchmark while Table 3 is on the 1,387 balanced split; the relationship and why the split was used should be stated when each table is introduced.
  2. [§3.2] Specify whether the VLM is shown the DCS statements verbatim and asked to judge their consistency with the image, or whether it must generate its own description to compare.
  3. [Figure 9] In Figure 9, indicate which B corresponds to the main results (or if results are averaged over B).
  4. [Abstract] The word 'significantly' in the abstract and conclusion is not supported by statistical tests; consider replacing with 'improves' and adding quantitative support.
  5. [Appendix C] The probe generation prompt includes 'gold answers' - describe how gold answers are derived (from atomic statements) and whether they were also human-validated beyond the 50-sample study.

Circularity Check

0 steps flagged

No significant circularity; shared VLM judge is a validity caveat, not a definitional reduction.

full rationale

The paper's central derivations are self-contained rather than circular. The OmniPhys benchmark probes are generated from a Physical Knowledge Graph anchored to PhET simulations and curriculum-aligned atomic physical statements, not from the optimizer's outputs; human evaluation independently validates probe correctness (96-100% validity, Gwet's AC1 0.892-0.986). OmniPrompt optimizes a meta-policy on a fixed 37-prompt training pool using DCS-only feedback from Gemini-2.5-Pro, while the final reported Joint Score (VQA x DCS) is computed on a held-out, PKP-balanced test split. The test score is not algebraically identical to the training reward, so the reported gains are not forced by construction. The main concern is external validity rather than circularity: the same VLM serves as both the training-time feedback signal and the final evaluation oracle, and the human evaluation reports only 80% human-VLM agreement with AC1=0.637 on image labeling, performed on non-optimized outputs. This is a measurement-risk caveat, not a reduction of the prediction to its inputs. The only self-citations (e.g., refs. [22] and [23], which include co-authors) are used for generic background concepts such as knowledge graphs and the term 'physical hallucinations'; they are not load-bearing and no uniqueness/ansatz is imported from them. Therefore the paper exhibits no significant circularity beyond a minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claims depend on explicit methodological choices—K=4, B tuned per backbone, a 37-prompt training pool, a reweighted split, and the GPT-4o/Gemini-2.5-Pro stack. None of these are derived from first principles. The most consequential assumptions are that a VLM can reliably judge physics from 2D images and that LLM-generated probes are correct; human validation covers only 50 probe pairs and 200 images.

free parameters (5)
  • Meta-policy P(t) = Optimized natural-language instruction; final text not published in paper
    The object of optimization (Eq. 6); fitted on a 37-prompt pool with validation rollback. The paper does not report the final policy verbatim.
  • Batch size B = Not stated for final results; ablation range 2–12
    Controls gradient aggregation in Eq. 4–6. Appendix B shows FLUX.1-dev is sensitive, with a sharp drop outside B∈[6,10].
  • Ensemble size K = 4
    Number of images per query used for evaluation and feedback (K=4 in §3.2 and §4.2.2); no sensitivity analysis for K is reported.
  • Training prompt pool size = 37 prompts
    ~2–3 prompts per PKP; chosen for computational cost rather than derived from any principle.
  • Diffraction-of-light reweighting = 224 to 60 prompts
    The PKP-balanced experimental split downsamples diffraction from 224 to 60, changing the aggregate score composition without a stated physical or psychometric justification.
axioms (6)
  • domain assumption A static 2D image is sufficient to unambiguously decide the binary physical outcomes for all 14 PKPs.
    §3.1.1 filters PKPs to observable binary states, but subtle states such as diffraction fringes and communicating-vessel equilibrium are hard to read from 2D renderings; the human study covered only 50 probes.
  • domain assumption Gemini-2.5-Pro's VLM judgments are a reliable proxy for human physical-consistency judgment.
    Table 4 gives human–VLM agreement of 80% and AC1=0.637 on 200 images; this assumption is load-bearing for every reported score.
  • domain assumption LLM-generated atomic physical statements and DCS probes are physically correct and unambiguous.
    Human validation reports 96–100% validity on 50 probe pairs, a small sample relative to the 1,551-prompt benchmark.
  • domain assumption TextGrad linguistic gradients provide a usable discrete analog of gradient descent for prompt-policy updates.
    §4.3 invokes 'textual gradient descent' without convergence or stability guarantees; only empirical trajectories support it.
  • domain assumption Aligning PhET simulations with standard curricula yields a sound ontology of physical knowledge points.
    §3.1 stage (1) relies on this alignment; no independent audit of the PhET-to-PKG mapping is provided beyond author/LLM curation.
  • domain assumption Prompt enhancement can improve physical alignment of a frozen T2I generator.
    Core premise of OmniPrompt, supported by Table 2's (ep) gains, but it also creates the leakage risk that expanded prompts state the expected outcome.

pith-pipeline@v1.3.0-alltime-deepseek · 18505 in / 14991 out tokens · 172194 ms · 2026-08-01T01:49:42.636631+00:00 · methodology

0 comments
read the original abstract

While text-to-image models exhibit remarkable visual fidelity, they frequently violate fundamental physical commonsense. Existing benchmarks often rely on coarse-grained descriptions, failing to diagnose the mastery of specific physical principles. Moreover, the high stochasticity of generative processes causes current prompt optimization methods to suffer from gradient hallucinations, where optimizers are misled by transient visual artifacts rather than systemic flaws. To address these challenges, we introduce OmniPhys, a rigorous benchmark of 1,551 samples grounded in a Physical Knowledge Graph. By aligning PhET simulations with standard curricula, OmniPhys operationalizes a knowledge-to-scenario pipeline that performs diagnostic stress tests via a dual-path verification protocol. We further propose OmniPrompt, an iterative framework that treats physical alignment as a discrete optimization problem. For each query, OmniPrompt aggregates K stochastic images into a per-query feedback buffer. Across training, it further merges feedback from batches of B queries before each meta-policy update, filtering seed and query-local noise. Evaluations across 12 representative text-to-image models reveal universal physical bottlenecks. Results demonstrate that OmniPrompt significantly enhances physical consistency across diverse backbones, proving the transferability and efficacy of our evolved meta-policies. The code and data are available at https://github.com/zjukg/OmniPhys

Figures

Figures reproduced from arXiv: 2607.25641 by Huajun Chen, Jeff Z. Pan, Jiaoyan Chen, Mingchen Tu, Wen Zhang, Yajing Xu, Yarong Lan, Yichi Zhang, Zhizhen Liu.

Figure 1
Figure 1. Figure 1: Examples of physical commonsense violations [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OmniPhys Infrastructure. (a) Construction: A hierarchical pipeline mapping PhET simulations to a PKG for synthesiz [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: OmniPhys Taxonomy. Hierarchical distribution of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The OmniPrompt Framework. An overview of the iterative alignment loop for physical commonsense in T2I generation. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Granular performance across 14 PKPs. Heatmaps show Joint Scores across Optics (left), Object Properties (center), and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Granular knowledge point recovery on the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance gains across physical domains. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-Model Policy Generalization. Results show [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sensitivity analysis and optimization dynamics. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 2 canonical work pages

  1. [1]

    Jingkun An, Yinghao Zhu, Zongjian Li, Enshen Zhou, Haoran Feng, Xijie Huang, Bohua Chen, Yemin Shi, and Chengwei Pan. 2025. AGFSync: Leveraging AI- Generated Feedback for Preference Optimization in Text-to-Image Generation. In AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, P...

    work page doi:10.1609/aaai.v39i2.32168 2025

  2. [2]

    Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, Yimeng Wang, Kai Yu, Wenxuan Chen, Ziwei Feng, Zijian Gong, Jianzhuang Pan, Yi Peng, Rui Tian, Siyu Wang, Bo Zhao, Ting Yao, and Tao Mei. 2025. HiDream-I1: A High-Efficient Image Genera- tive Foundation Model with Sparse Diffusion Transf...

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.22705 2025

  3. [3]

    Kaijie Chen, Zihao Lin, Zhiyang Xu, Ying Shen, Yuguang Yao, Joy Rimchala, Jiaxin Zhang, and Lifu Huang. 2025. R2I-Bench: Benchmarking Reasoning-Driven Text-to-Image Generation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025, Christos Christodoulopoulos, Tanmoy Chakrab...

    work page doi:10.18653/v1/2025.emnlp-main.636 2025

  4. [4]

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. 2025. Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling.CoRRabs/2501.17811 (2025). arXiv:2501.17811 doi:10.48550/ARXIV.2501.17811

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.17811 2025

  5. [5]

    Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xi- aoxia Hou, Weilin Huang, Yixuan Huang, Xiaowen Jian, Huafeng Kuang, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, Wei Liu, Yanzuo Lu, Zhengxiong Luo, Tongtong Ou, Guang Shi, Yichun Shi, Shiqi Sun, Yu Tian, Zhi Tian, Peng Wang, Rui Wang, Xun Wang, Ye Wang, Gu...

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.20427 2025

  6. [6]

    Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jinsheng Wang, Wenxuan Wang, Yueze Wang, Chengyuan Wang, Fan Zhang, Yingli Zhao, Ting Pan, Xianduo Li, Zecheng Hao, Wenxuan Ma, Zhuo Chen, Yulong Ao, Tiejun Huang, Zhongyuan Wang, and Xinlong Wang. 2025. Emu3.5: Native Multimodal Models are World Learners. CoRR...

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.26583 2025

  7. [7]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. 2024. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. InForty-first International Conference on Machine Learning, ICML 2024,...

    2024

  8. [8]

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. 2023. DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models. CoRRabs/2305.16381 (2023). arXiv:2305.16381 doi:10.48550/ARXIV.2305.16381

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.16381 2023

  9. [9]

    Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang

    Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun R. Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. 2023. Training- Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net....

    2023

  10. [10]

    Xingyu Fu, Muyu He, Yujie Lu, William Yang Wang, and Dan Roth. 2024. Commonsense-T2I Challenge: Can Text-to-Image Generation Models Under- stand Commonsense?CoRRabs/2406.07546 (2024). arXiv:2406.07546 doi:10. 48550/ARXIV.2406.07546

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.07546 2024

  11. [11]

    Google DeepMind. 2025. Gemini 3 Pro Image (Nano Banana Pro). https://ai. google.dev/gemini-api/docs/gemini-3. Accessed: 2026-02-09

    2025

  12. [12]

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. 2025. Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025. Computer Vision Foundation / IEEE, 15733–15744. doi...

    arXiv 2025

  13. [13]

    Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. 2023. Optimizing Prompts for Text-to-Image Generation. InAdvances in Neural Information Process- ing Systems 36: Annual Conference on Neural Information Processing Sys- tems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and...

    2023

  14. [14]

    Nailei Hei, Qianyu Guo, Zihao Wang, Yan Wang, Haofen Wang, and Wenqiang Zhang. 2024. A User-Friendly Framework for Generating Model-Preferred Prompts in Text-to-Image Synthesis. InThirty-Eighth AAAI Conference on Artifi- cial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Sympo...

    2024

  15. [15]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. InEMNLP (1). Association for Computational Linguistics, 7514–7528

    2021

  16. [16]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. InNIPS. 6626–6637

    2017

  17. [17]

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. 2023. T2I-CompBench: A Comprehensive Benchmark for Open-world Composi- tional Text-to-image Generation. InAdvances in Neural Information Process- ing Systems 36: Annual Conference on Neural Information Processing Sys- tems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, Alice ...

    2023

  18. [18]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. 2025. FLUX.1 Kontext: Flow Matching for In-Context ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.15742 2025

  19. [19]

    Fanqing Meng, Wenqi Shao, Lixin Luo, Yahong Wang, Yiran Chen, Quanfeng Lu, Yue Yang, Tianshuo Yang, Kaipeng Zhang, Yu Qiao, and Ping Luo. 2024. PhyBench: A Physical Commonsense Benchmark for Evaluating Text-to-Image Models.CoRR abs/2406.11802 (2024). arXiv:2406.11802 doi:10.48550/ARXIV.2406.11802

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.11802 2024

  20. [20]

    Yuwei Niu, Munan Ning, Mengren Zheng, Bin Lin, Peng Jin, Jiaqi Liao, Kun- Peng Ning, Bin Zhu, and Li Yuan. 2025. WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation.CoRRabs/2503.07265 (2025). arXiv:2503.07265 doi:10.48550/ARXIV.2503.07265

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.07265 2025

  21. [21]

    OpenAI. 2024. GPT-4o API Documentation (Snapshot: 2024-11-20). https:// platform.openai.com/docs/models/gpt-4o

    2024

  22. [22]

    J.Z. Pan, G. Vetere, J.M. Gomez-Perez, and H. Wu (Eds.). 2017.Exploiting Linked Data and Knowledge Graphs for Large Organisations. Springer

    2017

  23. [23]

    Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russa Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. 2023. Large Language Models and Knowledge Graphs: Opportunities and Challenges.Transactions on Graph...

    2023

  24. [24]

    Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng

  25. [25]

    Qi Qin, Le Zhuo, Yi Xin, Ruoyi Du, Zhen Li, Bin Fu, Yiting Lu, Jiakang Yuan, Xinyue Li, Dongyang Liu, Xiangyang Zhu, Manyuan Zhang, Will Beddow, Er- wann Millon, Victor Perez, Wenhai Wang, Conghui He, Bo Zhang, Xiaohong Liu, Hongsheng Li, Yu Qiao, Chang Xu, and Peng Gao. 2025. Lumina-Image 2.0: A Unified and Efficient Image Generative Framework.CoRRabs/25...

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.21758 2025

  26. [26]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 10674–10685. doi:10.1109/CVPR52688. 2022.01042

    arXiv 2022

  27. [27]

    Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mo- hammad Norouzi. 2022. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. InAdvances in Neural Information Process- ing...

    2022

  28. [28]

    Gemini Team. 2025. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities.CoRR abs/2507.06261 (2025). arXiv:2507.06261 doi:10.48550/ARXIV.2507.06261

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.06261 2025

  29. [29]

    Z.-Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven C. H. Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Junhan Shi, Qilong Wu, Feng Yu, Chi Zhang, Shifeng Zhang, and Shilin Zhou. 2025. Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion T...

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.22699 2025

  30. [30]

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Xiaofeng Meng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wa...

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.20314 2026

  31. [31]

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng- ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingku...

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.02324 2025

  32. [32]

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2023. ImageReward: Learning and Evaluating Hu- man Preferences for Text-to-Image Generation. InAdvances in Neural Infor- mation Processing Systems 36: Annual Conference on Neural Information Pro- cessing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, Decembe...

    2023

  33. [33]

    Cilin Yan, Jingyun Wang, Lin Zhang, Ruihui Zhao, Xiaopu Wu, Kai Xiong, Qing- song Liu, Guoliang Kang, and Yangyang Kang. 2025. Efficient and Accurate Prompt Optimization: the Benefit of Memory in Exemplar-Guided Reflection. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, Vienna, ...

    2025

  34. [34]

    Le, Denny Zhou, and Xinyun Chen

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. 2024. Large Language Models as Optimizers. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net. https://openreview.net/forum?id= Bb4VGOWELI

    2024

  35. [35]

    ex- panded_text

    Daoan Zhang, Che Jiang, Ruoshi Xu, Biaoxiang Chen, Zijian Jin, Yutian Lu, Jianguo Zhang, Liang Yong, Jiebo Luo, and Shengda Luo. 2025. WorldGenBench: A World-Knowledge-Integrated Benchmark for Reasoning-Driven Text-to-Image Generation.CoRRabs/2505.01490 (2025). arXiv:2505.01490 doi:10.48550/ARXIV. 2505.01490 A Implementation of System Prompt Initializatio...

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.01490 2025

  36. [2023]

    Gradient Descent

    Automatic Prompt Optimization with "Gradient Descent" and Beam Search. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, 7957–7968. doi:10.18653/V1/2023.EMNLP-MAIN.494

    doi:10.18653/v1/2023.emnlp-main.494 2023