Pith. sign in

REVIEW 4 major objections 5 minor 6 cited by

Maestro claims that a fully autonomous, test-time loop of multimodal agents can improve text-to-image generation from a single initial prompt, outperforming both the original prompt and existing automated prompt-optimization methods on hard

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

2026-08-04 17:35 UTC pith:SWKFJJS7

load-bearing objection Maestro is a solid test-time T2I optimizer, but the DVQ-as-ground-truth loop makes the main evaluation partly circular; worth refereeing with conditions. the 4 major comments →

arxiv 2509.10704 v1 pith:SWKFJJS7 submitted 2025-09-12 cs.AI cs.CV

Maestro: Self-Improving Text-to-Image Generation via Agent Orchestration

Xingchen Wan , Han Zhou , Ruoxi Sun , Hootan Nakhost , Ke Jiang , Rajarishi Sinha , Sercan \"O. Ar{\i}k This is my paper
classification cs.AI cs.CV
keywords text-to-image generationprompt optimizationmultimodal LLM agentsself-critiquepairwise comparisonvisual question answeringtest-time optimizationMLLM-as-a-judge
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.

Maestro is a test-time system for text-to-image models that needs only the user's initial prompt. It decomposes that prompt into a list of visual questions, then runs an iterative loop: generate an image, have critic agents explain which visual questions it fails and how to fix them, ask an editor to propose prompt revisions, and use a separate judge to compare new images head-to-head against the best so far. The paper claims this loop outperforms the original prompt, a one-shot LLM rewrite, a trained prompt optimizer (Promptist), and two prior test-time optimizers (LM-BBO, OPT2I) on two hard benchmarks, with a budget of eight image-generation calls. The authors argue the pairwise comparison objective avoids the distortions of scalar reward models, and that the system improves as stronger multimodal models are used as critics and judges.

Core claim

The central claim is that prompt engineering for black-box text-to-image models can be fully automated by combining structured visual-question decomposition with generative, interpretable critique and pairwise preference tracking. At initialization, an LLM rewrites the user prompt following best practices and decomposes it into decomposed visual questions (DVQs), a list of yes/no questions that each probe a desired property. During each iteration, the system generates an image, scores it against the DVQs, and extracts textual rationales for any 'No' answers; these rationales drive targeted prompt edits, while a second, holistic improvement path proposes broader revisions. A verifier checks e

What carries the argument

The load-bearing mechanism is the decomposed visual question (DVQ) set: a list of yes/no questions, generated once from the user prompt, that serves as a structured, interpretable proxy for user intent. DVQs drive three things: the critic agents (which identify which DVQs the current best image fails and produce textual rationales and edit suggestions), the targeted-editing generator (which turns those rationales into prompt revisions), and the verifier (which re-checks revised prompts against the DVQs). The second key mechanism is the pairwise comparator: instead of optimizing a scalar score, an MLLM judge conducts binary tournaments between the new image and the incumbent best, with repeat

Load-bearing premise

The single load-bearing premise is that the decomposed visual questions, generated once from the user prompt, faithfully and completely capture what the user actually wants; if they don't, the critics, editor, and verifier will all confidently optimize toward the wrong target.

What would settle it

Take a prompt whose intent is deliberately hard to decompose into yes/no questions (e.g., a specific mood or artistic ambiguity) and run Maestro against a human-in-the-loop baseline; if human raters reliably prefer the baseline's images, or if Maestro returns images that satisfy all DVQs but miss the stated intent, the central claim fails. A sharper test: corrupt the DVQs (leave out one critical question) and show that Maestro never recovers the missing attribute, because the verifier enforces the omission.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the claimed results hold, users can hand an underspecified prompt to Maestro and receive a better prompt-image pair without any manual iteration.
  • The method is training-free and black-box, so it can wrap any T2I API and immediately benefit from newer, stronger MLLMs as critics and judges.
  • The pairwise comparison objective sidesteps the need for a calibrated scalar reward, making the approach applicable to other generative tasks where ground truth is absent, such as audio or video synthesis.
  • The initial LLM rewrite step means train-time prompt optimizers can be plugged in as initialization, combining the strengths of both paradigms.
  • The returned optimized prompt can be inspected and reused, since the system returns both prompt and image, giving interpretability to the improvement process.

Where Pith is reading between the lines

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

  • The DVQ decomposition is a hidden control knob: if the generated questions miss a user's real intent (e.g., mood, aesthetics, abstract concepts), the critics will never address it and the verifier will actively block attempts to add it — a testable failure mode by comparing DVQs against human-written intent lists.
  • The judge and the editor are typically the same family of MLLM, so there is a risk of self-preference bias where the judge systematically favors images produced by its own edits; a cross-model test (optimizer from model X, judge from model Y) would quantify this.
  • Because the system evolves both prompt and image, the improved prompts could be harvested as training data for train-time prompt optimizers, compounding gains over successive deployments.
  • The pairwise tournament with eight T2I calls suggests a natural scaling law: more budget or stronger judges should push quality further, but also raises the question of diminishing returns and cost ceilings for production use.

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 proposes Maestro, a test-time, model-agnostic multi-agent system for improving text-to-image (T2I) generation. Given a user prompt, Maestro first generates a set of decomposed visual questions (DVQs) and an initial rewritten prompt. It then iterates: generate an image, answer the DVQs with an MLLM, maintain a best prompt/image via pairwise MLLM-judge comparisons, propose new prompts through both targeted editing of DVQ failures and implicit holistic improvement, and finally verify/self-correct the prompt against the same DVQs. The method is evaluated on p2-hard and DSG-1K using Imagen 3, comparing against Original, Rewrite, Promptist, LM-BBO, and OPT2I. Results include DSGScore improvements, AutoSxS pairwise win rates, a small human study against LM-BBO, ablations showing each component helps, and scaling with stronger MLLM backbones.

Significance. If the results hold, Maestro offers a practical and interpretable way to improve black-box T2I generation using only test-time compute, with no training or labeling. The paper includes several strengths: a clear dual-generator design, a pairwise objective that sidesteps scalar reward issues, a verifier module explicitly designed to prevent semantic drift, and experiments that cover multiple baselines, datasets, judge models, and an ablation study. The AutoSxS evaluation with a held-out judge (Gemini 2.0 Flash) and the agreement with a human study are positive signs. However, the central claim that Maestro improves fidelity to user intent depends on an unvalidated DVQ decomposition that is reused both as the optimization signal and as the evaluation metric; this circularity and the limited human validation are the main risks. The contribution is primarily an orchestration of existing components (DSG-style decomposition, MLLM-as-judge, iterative LLM editing), but the combination is effective and the empirical story is fairly compelling.

major comments (4)
  1. [§3.1, §3.3, §3.4, Table 1, App. A] The DVQ fidelity/circularity concern is load-bearing. The same DVQs generated once from the user prompt are used in three places: targeted editing (§3.3) selects failures based on DVQ 'No' answers, Verify-and-Self-Correct (§3.4) checks new prompts against the same DVQs, and the DSGScore evaluation in Table 1 is itself a DVQ-based metric. App. A states that Maestro's DVQs reuse the instructions and few-shot examples from Cho et al. (2024). Consequently, a systematic error in the DVQ decomposition (omitted requirement, wrong emphasis, or mis-parsed constraint) is inherited by both the optimizer and the evaluator. The system can converge to an image that satisfies the distorted DVQ checklist while diverging from the user's actual intent, and the self-verifier cannot detect this because it checks against the same fallible DVQs. The AutoSxS evaluation is a partial safeguard because the judge
  2. [§4, 'Datasets and models'] The dataset filtering step retains only prompts for which the original prompt never reaches DSGScore 1.0 in 8 attempts. This conditions the evaluation set on the DSGScore metric that is later used as a headline result, and it may systematically select for prompts where the original prompt is poorly aligned with DSG questions, potentially inflating the observed gains. While focusing on challenging cases is defensible, the filtering should be justified independently, e.g., by also reporting results on the unfiltered set or by using a different criterion (e.g., human difficulty ratings) for filtering.
  3. [Fig. 6 and App. A 'Human preference studies'] The human evaluation is very small (3 respondents, all described as the authors' colleagues) and compares Maestro only against LM-BBO, not against the original prompt. Given the central claim is that Maestro improves fidelity to user intent, the absence of a human comparison against the original prompt is a notable gap. At minimum, a larger human study with independent raters and a condition comparing Maestro's final image to the image from the original prompt is needed to support the 'faithfulness' component of the claim.
  4. [Table 1 and Fig. 5] The paper reports no significance tests or confidence intervals for the DSGScore differences or the AutoSxS win rates. For example, Maestro vs. OPT2I DSGScore is 0.921 ± 0.10 vs. 0.900 ± 0.09 on p2-hard; with the reported standard deviations, the difference may not be statistically significant. Similarly, the AutoSxS win/lose distributions in Fig. 5 are not accompanied by paired tests. Because several claims are comparative ('consistently outperforms', 'significantly improves'), paired bootstrap or Wilcoxon tests should be reported for both DSGScore and AutoSxS advantage.
minor comments (5)
  1. [Throughout] The text contains several formatting artifacts: 'M ae s t ro' with irregular spacing in the abstract and body; 'generatd' typo in App. A; Table 3 has duplicate numbering ('6.' appears twice).
  2. [§2, train-time approaches] The sentence 'Other works like have expanded in this direction' is missing a citation/name; please fix.
  3. [Fig. 2 caption] The caption says 'left: pu; right: p(0)' but the images are not clearly labeled in the text; consider adding explicit (a)/(b) labels in the figure.
  4. [§3.2, pairwise comparison] The judge queries 2n times with n=3, but the temperature and tie-breaking details are mentioned only later in the text; consider stating these in the main procedure for clarity.
  5. [Table 1] The rank metric is introduced but its interpretation is not explained in the caption; please clarify how rank is computed and what values are better (the caption says lower is better, which is clear, but the scale is not).

Circularity Check

1 steps flagged

DSGScore gains are partly self-referential because targeted editing and the DSGScore evaluation share the same self-generated DVQ checklist; the AutoSxS and human studies keep the core claim substantially independent.

specific steps
  1. other [§3.1 (DVQ generation), §3.3 (Targeted editing), §3.4 (Verify & Self-correct), Table 1, Appendix A]
    "this generator reuses the DVQs generated at initialization 𝑄 and the MLLM’s responses to them for the best image so far 𝑅∗:𝑡−1. ... we 1) identify a subset of DVQs where the MLLM judge provided a 'No' answer ... 3) conditional on the rationalization, a specific edit or rewrite to the prompt aimed precisely at rectifying the identified shortcoming ... These DVQs were generated once, and are subsequently fixed for all methods that use these questions (e.g., Maestro and OPT2I)."

    Maestro's targeted editing and self-verification explicitly optimize and constrain prompts against the DVQs generated once from the user prompt in §3.1. Table 1 reports DSGScore as an evaluation metric, and Appendix A confirms DSGScore uses the same fixed DVQ set for all methods. Thus the DSGScore improvement in Table 1 partly measures how well Maestro satisfies its own self-generated checklist, rather than an independently established notion of user intent. This makes the DSGScore comparison partially circular, although the AutoSxS (using Gemini 2.0 Flash, a judge not used in optimization) and the human study remain external validation.

full rationale

The paper's central pipeline is not circular in the narrow sense: Maestro does not fit parameters and then predict them; its comparator is a pairwise MLLM judge conditioned on the user prompt, and the headline AutoSxS evaluation uses a held-out judge (Gemini 2.0 Flash) different from the optimizer (Gemini 1.5 Pro). The human side-by-side study is also external to the DVQ machinery. However, one reported result is partially self-referential: the DSGScore in Table 1 is computed from the same DVQ set that Maestro's targeted editing (§3.3) and Verify-and-Self-Correct (§3.4) explicitly optimize against. Because the DVQs are generated once from the user prompt and fixed for all methods, the DSGScore comparison is partly a comparison of how well each method satisfies that particular LLM-generated checklist. This does not invalidate the paper's main claim — the AutoSxS and human evaluations are substantially independent — but it lowers the evidentiary weight of Table 1. The self-citations (e.g., Zhou et al. 2024a, 2024b) are for position-bias mitigation and calibration; they are standard and not load-bearing. The additional concern that a bad initial DVQ decomposition cannot be corrected by the verifier is a validity/robustness issue, not a circular-derivation issue, and is not counted heavily here.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The system is an engineered pipeline; it introduces no fitted physical constants or new entities. Its load-bearing assumptions are domain assumptions about the fidelity of DVQ decomposition and the reliability of MLLM judges, plus hand-set hyperparameters that are not fitted to data.

free parameters (5)
  • DVQ 'No' threshold 0.5
    Responses are binarized at p(Yes|image, question) < 0.5 to select which DVQs trigger targeted editing (Sec. 3.3). Hand-set, not fit, but determines which failures are edited.
  • judge query count n=3
    Number of duplicate pairwise judge queries with swapped positions to mitigate position bias (Sec. 3.2 and App. A). Hand-set.
  • judge temperature 0.7
    Sampling temperature for judge model queries (Sec. 3.2). Hand-set.
  • self-verification patience 3
    Patience for the verify-and-self-correct loop (Sec. 3.4). Hand-set.
  • max T2I calls 8
    Budget cap used for all compared methods (Sec. 4). Hand-set.
axioms (3)
  • domain assumption DVQs generated once from the user prompt fully capture the user's intent
    All of targeted editing, verification, and the DSGScore evaluation rest on the DVQ decomposition from Sec. 3.1. If DVQs omit or misstate a key requirement, the system will confidently optimize toward a distorted objective.
  • domain assumption MLLM judge pairwise preferences approximate human preferences well enough to guide self-improvement
    The comparator (Sec. 3.2) and AutoSxS rely on Gemini models as judges; App. B.1 shows judge choice changes tie rates, and the human study is the only external check, with only 3 respondents.
  • domain assumption Posthoc rationalization of 'No' VQA answers yields useful textual edit signals
    Targeted editing (Sec. 3.3) assumes the rationalizations and suggestions generated by the MLLM are actionable and improve the prompt rather than introducing noise; no independent verification of rationale quality is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 20836 in / 11833 out tokens · 123478 ms · 2026-08-04T17:35:12.230696+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Maestro: Self-Improving Text-to-Image Generation via Agent Orchestration." pith.science (2026). https://pith.science/paper/SWKFJJS7

@misc{pith2026250910704,
  author       = {Pith},
  title        = {Pith review of: Maestro: Self-Improving Text-to-Image Generation via Agent Orchestration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWKFJJS7}},
  note         = {Machine review of arXiv:2509.10704}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-to-image (T2I) models, while offering immense creative potential, are highly reliant on human intervention, posing significant usability challenges that often necessitate manual, iterative prompt engineering over often underspecified prompts. This paper introduces Maestro, a novel self-evolving image generation system that enables T2I models to autonomously self-improve generated images through iterative evolution of prompts, using only an initial prompt. Maestro incorporates two key innovations: 1) self-critique, where specialized multimodal LLM (MLLM) agents act as 'critics' to identify weaknesses in generated images, correct for under-specification, and provide interpretable edit signals, which are then integrated by a 'verifier' agent while preserving user intent; and 2) self-evolution, utilizing MLLM-as-a-judge for head-to-head comparisons between iteratively generated images, eschewing problematic images, and evolving creative prompt candidates that align with user intents. Extensive experiments on complex T2I tasks using black-box models demonstrate that Maestro significantly improves image quality over initial prompts and state-of-the-art automated methods, with effectiveness scaling with more advanced MLLM components. This work presents a robust, interpretable, and effective pathway towards self-improving T2I generation.

discussion (0)

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

Forward citations

Cited by 6 Pith papers

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

  1. Semantic Browsing: Controllable Diversity for Image Generation

    cs.CV 2026-06 unverdicted novelty 7.0

    A technique for controllable diversity in text-to-image generation by inducing structured semantic variations at the prompt level via VLM and agentic workflow.

  2. GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

    cs.CV 2026-05 unverdicted novelty 7.0

    GenEvolve proposes a self-evolving agent framework for open-ended image generation that uses tool-orchestrated trajectories and visual experience distillation from best-worst differences to achieve reported state-of-t...

  3. Reasoning Arena: Trace Tournaments When Verifiable Rewards Fall Short

    cs.LG 2026-06 unverdicted novelty 6.0

    Reasoning Arena converts non-diverse reward groups in RLVR into relative rewards via adaptive trace tournaments and Bradley-Terry fitting on anchor comparisons, claiming 7.6% average gains and 27-41% faster training o...

  4. Visual Persuasion: What Influences Decisions of Vision-Language Models?

    cs.CV 2026-02 conditional novelty 6.0

    Naturalistic edits to background, lighting, and staging systematically shift the choices of frontier vision-language models, and a new optimization-plus-interpretation framework (CVPO) discovers and explains the visua...

  5. APEX: Automated Prompt Engineering eXpert with Dynamic Data Selection

    cs.CL 2026-06 unverdicted novelty 5.0

    APEX dynamically tiers data into Easy/Hard/Mixed based on optimization lineage and prioritizes Mixed examples, reporting 11.2% and 6.8% average gains over baseline prompts on two models under a 5,000-call budget.

  6. GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

    cs.CV 2026-05 unverdicted novelty 5.0

    GenEvolve introduces a self-evolving agent framework for image generation using tool-orchestrated trajectories and Visual Experience Distillation to achieve claimed SOTA results on benchmarks.

Reference graph

Works this paper leans on

49 extracted references · 18 linked inside Pith · cited by 5 Pith papers

  1. [1]

    Baldridge, J

    J. Baldridge, J. Bauer, M. Bhutani, N. Brichtova, A. Bunner, L. Castrejon, K. Chan, Y. Chen, S. Dieleman, Y. Du, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024

    arXiv 2024

  2. [2]

    Betker, G

    J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2 0 (3): 0 8, 2023

    2023

  3. [3]

    R. A. Bradley and M. E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39 0 (3/4): 0 324--345, 1952

    1952

  4. [4]

    T. Cao, C. Wang, B. Liu, Z. Wu, J. Zhu, and J. Huang. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Chiang and H.-y

    C.-H. Chiang and H.-y. Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    J. Cho, Y. Hu, R. Garg, P. Anderson, R. Krishna, J. Baldridge, M. Bansal, J. Pont-Tuset, and S. Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation. International Conference on Learning Representations, 2024

    2024

  7. [7]

    P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017

    2017

  8. [8]

    Datta, A

    S. Datta, A. Ku, D. Ramachandran, and P. Anderson. Prompt expansion for adaptive text-to-image generation. arXiv preprint arXiv:2312.16720, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Delmas, P

    G. Delmas, P. Weinzaepfel, T. Lucas, F. Moreno-Noguer, and G. Rogez. Posescript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346--362. Springer, 2022

    2022

  10. [10]

    Y. Fan, O. Watkins, Y. Du, H. Liu, M. Ryu, C. Boutilier, P. Abbeel, M. Ghavamzadeh, K. Lee, and K. Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36: 0 79858--79885, 2023

    2023

  11. [11]

    M. Hahn, W. Zeng, N. Kannen, R. Galt, K. Badola, B. Kim, and Z. Wang. Proactive agents for multi-turn text-to-image generation under uncertainty. arXiv preprint arXiv:2412.06771, 2024

    arXiv 2024

  12. [12]

    Y. Hao, Z. Chi, L. Dong, and F. Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36: 0 66923--66939, 2023

    2023

  13. [13]

    Hessel, A

    J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y. Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021

    Pith/arXiv arXiv 2021

  14. [14]

    Y. Hu, B. Liu, J. Kasai, Y. Wang, M. Ostendorf, R. Krishna, and N. A. Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406--20417, 2023

    2023

  15. [15]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Kirstain, A

    Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 0 36652--36663, 2023

    2023

  17. [17]

    H. Lee, S. Phatale, H. Mansoor, T. Mesnard, J. Ferret, K. Lu, C. Bishop, E. Hall, V. Carbune, A. Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. International Conference on Machine Learning, 2024

    2024

  18. [18]

    S. Liu, S. Yu, Z. Lin, D. Pathak, and D. Ramanan. Language models as black-box optimizers for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12687--12697, 2024

    2024

  19. [19]

    Mahan, D

    D. Mahan, D. Van Phung, R. Rafailov, C. Blagden, N. Lile, L. Castricato, J.-P. Fr \"a nken, C. Finn, and A. Albalak. Generative reward models. arXiv preprint arXiv:2410.12832, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Ma \ n as, P

    O. Ma \ n as, P. Astolfi, M. Hall, C. Ross, J. Urbanek, A. Williams, A. Agrawal, A. Romero-Soriano, and M. Drozdzal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    W. Mo, T. Zhang, Y. Bai, B. Su, J.-R. Wen, and Q. Yang. Dynamic prompt optimizing for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26627--26636, 2024

    2024

  22. [22]

    Y. Niu, M. Ning, M. Zheng, B. Lin, P. Jin, J. Liao, K. Ning, B. Zhu, and L. Yuan. Wise: A world knowledge-informed semantic evaluation for text-to-image generation. arXiv preprint arXiv:2503.07265, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

    2022

  24. [24]

    gradient descent

    R. Pryzant, D. Iter, J. Li, Y. T. Lee, C. Zhu, and M. Zeng. Automatic prompt optimization with" gradient descent" and beam search. Empirical Methods in Natural Language Processing, 2023

    2023

  25. [25]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 0 (2): 0 3, 2022

    Pith/arXiv arXiv 2022

  26. [26]

    M. Reid, N. Savinov, D. Teplyashin, D. Lepikhin, T. Lillicrap, J.-b. Alayrac, R. Soricut, A. Lazaridou, O. Firat, J. Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    J. Ren, Y. Zhao, T. Vu, P. J. Liu, and B. Lakshminarayanan. Self-evaluation improves selective generation in large language models. In Proceedings on, pages 49--64. PMLR, 2023

    2023

  28. [28]

    C. Ross, M. Hall, A. R. Soriano, and A. Williams. What makes a good metric? evaluating automatic metrics for text-to-image consistency. arXiv preprint arXiv:2412.13989, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35: 0 36479--36494, 2022

    2022

  30. [30]

    L. Shi, C. Ma, W. Liang, W. Ma, and S. Vosoughi. Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms. arXiv preprint arXiv:2406.07791, 2024

    arXiv 2024

  31. [31]

    Turc and G

    I. Turc and G. Nemade. Midjourney user prompts & generated images (250k). Retrieved March, 15: 0 2024, 2022

    2024

  32. [32]

    R. Wang, T. Liu, C.-J. Hsieh, and B. Gong. On discrete prompt optimization for diffusion models. arXiv preprint arXiv:2407.01606, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Z. J. Wang, E. Montoya, D. Munechika, H. Yang, B. Hoover, and D. H. Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022

    Pith/arXiv arXiv 2022

  34. [34]

    J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 0 15903--15935, 2023

    2023

  35. [35]

    C. Yang, X. Wang, Y. Lu, H. Liu, Q. V. Le, D. Zhou, and X. Chen. Large language models as optimizers. In International Conference on Learning Representations, 2024

    2024

  36. [36]

    Yarom, Y

    M. Yarom, Y. Bitton, S. Changpinyo, R. Aharoni, J. Herzig, O. Lang, E. Ofek, and I. Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36: 0 1601--1619, 2023

    2023

  37. [37]

    J. Yu, Y. Xu, J. Y. Koh, T. Luong, G. Baid, Z. Wang, V. Vasudevan, A. Ku, Y. Yang, B. K. Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2 0 (3): 0 5, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    differentiation

    M. Yuksekgonul, F. Bianchi, J. Boen, S. Liu, Z. Huang, C. Guestrin, and J. Zou. Textgrad: Automatic" differentiation" via text. arXiv preprint arXiv:2406.07496, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    T. Yun, D. Zhang, J. Park, and L. Pan. Learning to sample effective and diverse prompts for text-to-image generation. arXiv preprint arXiv:2502.11477, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Z. Zhao, E. Wallace, S. Feng, D. Klein, and S. Singh. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , pages 12697--12706, 2021

    2021

  41. [41]

    Zheng, H

    C. Zheng, H. Zhou, F. Meng, J. Zhou, and M. Huang. Large language models are not robust multiple choice selectors. arXiv preprint arXiv:2309.03882, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    H. Zhou, X. Wan, Y. Liu, N. Collier, I. Vuli \'c , and A. Korhonen. Fairer preferences elicit improved human-aligned large language model judgments. In Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1241--1252, Miami, Florida, USA, Nov. 2024 a . Association f...

    2024

  43. [43]

    H. Zhou, X. Wan, L. Proleev, D. Mincu, J. Chen, K. A. Heller, and S. Roy. Batch calibration: Rethinking calibration for in-context learning and prompt engineering. In The Twelfth International Conference on Learning Representations, 2024 b

    2024

  44. [44]

    Y. Zhou, A. I. Muresanu, Z. Han, K. Paster, S. Pitis, H. Chan, and J. Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2022

    2022

  45. [45]

    D. M. Ziegler, N. Stiennon, J. Wu, T. B. Brown, A. Radford, D. Amodei, P. Christiano, and G. Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019

    Pith/arXiv arXiv 1909

  46. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  47. [47]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  48. [48]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  49. [49]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...