Pith. sign in

REVIEW 3 major objections 9 minor 45 references

PRISM teaches one vision-language model to rewrite text-to-image prompts by diagnosing the images those prompts produce.

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 · grok-4.5

2026-07-31 17:21 UTC pith:GSHM7C4I

load-bearing objection Solid incremental systems paper: unified rewrite+diagnose VLM with hybrid self-reward; Table 1 preference gains are partly circular, but TIFA/CompBench still look real. the 3 major comments →

arxiv 2607.24353 v1 pith:GSHM7C4I submitted 2026-07-27 cs.CV

PRISM: Prompt Refinement via Image-grounded Self-rewarding Mechanism for Text-to-Image Generation

Guo Tang , HongJie Luo , Tianxu Wang , Ying Zhang , Hao Wang This is my paper
classification cs.CV
keywords text-to-image generationprompt optimizationvision-language modelsself-rewarding learningvisual feedbacksemantic consistencymulti-objective rewardgroup sequence policy optimization
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.

Text-to-image models are brittle to how a prompt is worded. Most automatic fixers either rewrite text in isolation or bolt on separate critics at test time, so they never learn a reusable policy grounded in what the picture actually shows. PRISM closes that loop: a single vision-language model both rewrites the prompt and judges the resulting image along three axes—semantic consistency, aesthetic quality, and human preference—then turns those scores into a training signal. After multi-task supervised warm-up, the same model improves its own rewriting policy through self-rewarding updates that push every dimension toward an ideal while punishing the weakest one. The result is better holistic image quality, stronger fine-grained semantic match to the original user intent, and readable diagnoses that explain which prompt changes to make next.

Core claim

A unified vision-language model can learn a reusable prompt-optimization policy by first being supervised on joint rewriting and structured visual feedback, then being further trained with self-generated multi-dimensional image rewards. When those rewards combine overall closeness to an ideal score vector with a penalty on the worst dimension, the optimized prompts improve both broad image quality metrics and fine-grained compositional faithfulness relative to the original user prompt.

What carries the argument

Image-grounded self-rewarding with hybrid ideal-point and Chebyshev aggregation: candidate prompts are rendered by a frozen generator, scored by a frozen copy of the same model on semantic, aesthetic, and preference axes, turned into a scalar that rewards progress toward the ideal vector while penalizing the weakest axis, and used to update the prompt policy with group sequence policy optimization.

Load-bearing premise

The same model that rewrites prompts can also serve as a stable, unbiased judge of its own outputs, even though its training labels and the reported metrics share overlapping preference and aesthetic scorers.

What would settle it

Hold the generator fixed, replace PRISM’s self-judge with an independent human or held-out preference panel on the same BeautifulPrompt and T2I-CompBench prompts, and check whether the reported gains in PickScore, ImageReward, TIFA, and compositional averages disappear or reverse.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Prompt optimization can be learned once as a reusable policy rather than re-run as a multi-module test-time loop for every image.
  • Structured visual diagnoses become actionable rewrite cues, so users can see which missing objects, relations, or style flaws the optimizer is fixing.
  • Balancing overall score improvement with worst-dimension penalty yields more even gains across alignment, aesthetics, and preference than single-objective rewards.
  • The same trained policy transfers across generators, as shown by consistent gains when the frozen backend changes from Stable Diffusion 3.5 to FLUX.2-klein.
  • Iterative feedback rounds can further lift hard samples without changing the underlying model weights.

Where Pith is reading between the lines

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

  • If self-judging VLMs become standard reward models, prompt optimizers may start gaming shared aesthetic and preference scorers unless evaluation panels are deliberately de-correlated from training labels.
  • The hybrid ideal/Chebyshev objective is a portable multi-objective template that could be reused for other controllable generation loops (video, 3D, layout) where one weak axis ruins the result.
  • Feedback-guided rewriting data that pairs under-specified prompts with controlled degradations may be as important as the RL stage for teaching the model what ‘missing detail’ looks like.
  • Adaptive early stopping on per-sample feedback scores could cut the inference cost of multi-round refinement while keeping most of the quality gain.

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

3 major / 9 minor

Summary. The paper proposes PRISM, a prompt-optimization framework for text-to-image generation in which a single Qwen3-VL-8B VLM is trained to act both as prompt rewriter and as visual judge. Stage I is multi-task SFT over four instruction types (rewriting, feedback assessment, score calibration, feedback-guided rewriting); Stage II runs self-rewarding GSPO where a frozen copy of the SFT VLM scores candidate generations along semantic consistency, aesthetic quality, and preference alignment, aggregated by a hybrid ideal-point + Chebyshev reward (Eqs. 14–20). The authors report gains on the BeautifulPrompt test set (Table 1), TIFA and T2I-CompBench (Table 2), ablations on feedback source and reward aggregation (Tables 3–4), and cross-generator replication on FLUX.2-klein (Tables 6–7), plus anonymized code.

Significance. If the results hold, this is a solid contribution to prompt optimization: the unified rewriter/judge design with structured diagnosis is more interpretable than scalar-reward RL baselines, the hybrid ideal-point/Chebyshev aggregation is cleanly ablated (Table 4), the feedback-source ablation (Table 3) usefully shows the self-judge is nearly competitive with an external VLM judge at equal inference cost, and the TIFA/T2I-CompBench gains (Table 2) constitute evidence from evaluators largely outside the training-reward family. The paper also reports hyperparameters, dataset sizes, and discretization procedures in unusual detail (Appendix A), ships code, and — commendably — computes all evaluation metrics against the original user prompt rather than the optimized one, which guards against the common failure of optimizing alignment to the rewritten text. These are genuine strengths. The main reservations are the overlap between reward scorers and Table 1 evaluation metrics, and the absence of any variance or significance reporting.

major comments (3)
  1. [§4.3, Table 1; Appendix A.4, Eqs. (24)–(26)] Table 1 vs. Eqs. (24)–(26) and Appendix A.4: four of the five Table 1 columns are functions the policy was directly trained to maximize. The SFT judge's preference target is defined as the average of normalized ImageReward, PickScore, and HPSv2 (Eq. 24), the aesthetic target is a normalized external aesthetic metric (A.4), the VLM is distilled to reproduce these discretized scores, and Stage II then optimizes GSPO against this judge (Eqs. 14–20). Table 1 then evaluates on exactly PickScore, HPSv2, ImageReward, and Aesthetic. Improvements over baselines never trained against these scorers (Promptist, VisualPrompter, GPT-4o) are therefore partially by construction and do not, on their own, support the abstract's claim of improved 'holistic image quality and human preference alignment.' Table 2 (TIFA, T2I-CompBench) and Table 3 (Ext-VLM ablation) mitigate but do not resolve this: TIFA/CompB
  2. [§4.3–4.4, Tables 1, 2, 4] No variance or significance reporting anywhere. Several load-bearing margins are small: in Table 1's text-image group, PRISM's HPSv2 of 0.305 vs. PRISM-SFT's 0.303 and GPT-4o's 0.283; PickScore 22.526 vs. Qwen3-VL-235B's 22.381; in Table 4, Aesthetic 6.318 (Hybrid) vs. 6.312 (Weighted Sum) — a 0.006 gap used to claim the hybrid reward is 'best across all metrics.' T2I decoding is stochastic and these numbers are single-run means over fixed seeds/prompts. Please report standard errors (over multiple generation seeds per prompt, or bootstrap over prompts) for Tables 1, 2, and 4 at minimum, and soften claims where deltas are within noise.
  3. [Appendix A.4; §3.1, Eq. (2)] Semantic-feedback reference choice (Appendix A.4, below Eq. (23)): when scoring the image of the raw prompt p = x_i, the 'reference prompt' against which DSG-style QA accuracy is computed is x̂_i, the high-quality community prompt — not the user's original prompt. The semantic reward therefore measures alignment with the optimizer's own target distribution rather than fidelity to user intent. This is a defensible design (it is what creates the 'optimization gap'), but it means the Stage-II semantic reward can push the policy toward community-prompt stylistic conventions ('8k, cinematic lighting, octane render') rather than toward the user's intent, and the qualitative examples (Fig. 4) show exactly this expansion pattern. The paper claims the method 'preserves the user intent' (§3.1); please discuss this tension explicitly, and ideally show an intent-preservation measurement (e.g., an LL
minor comments (9)
  1. [Title page] Correspondence email is 'guot3907@gamil.com' — likely a typo for gmail.com.
  2. [Appendix A] Appendix A opening says 'We construct four types of training data' but then lists only three items (i)–(iii), and names D_vf whereas the main text and Table 5 use D_fb. Please reconcile notation.
  3. [§4.4, Figure 3] Figure 3 is referenced as showing iterative-refinement gains 'especially for samples with low first-pass scores,' but the caption gives no axis definitions, metric, or numeric values; as reproduced it is not independently interpretable.
  4. [Appendix A.6, Table 5] Table 5 mentions 'The size M depends on...' but M is never defined; the table already lists explicit sizes.
  5. [§4.3, Tables 1–2] Tables 1–2 mix methods with very different inference-time compute (single-pass rewriting vs. multi-round test-time refinement like VisualPrompter/TIR/OPT2I). Table 3's C_T2I/C_VLM accounting is good; please add the same per-input call counts (or wall-clock) to Tables 1–2 so the main comparison is cost-aware.
  6. [§4.1] Only two self-rewarding iterations are performed (§4.1). A brief note on whether more iterations help, saturate, or cause judge-gaming drift would strengthen the self-rewarding claim, especially given the frozen-twin judge of Eq. (14).
  7. [Figures 1–2; abstract] Figure 2 contains small wording issues ('worse-case correction,' inconsistent roman numerals 'Stage Ⅱ'), and the abstract hyphenation 'PRISM, a Prompt Refinement framework' is missing non-breaking spaces throughout ('we proposePRISM').
  8. [§4.2; Appendix A.2] TheBeautifulPrompt test set is used for evaluation while D_rew is drawn from its training pool — please state explicitly that the 2k test prompts are disjoint from the 10k SFT subset, and whether any near-duplicates were filtered.
  9. [§2.2] Consider citing and contrasting with reward-hacking analyses of preference-model-based T2I optimization (e.g., work on overoptimization of ImageReward/PickScore) — directly relevant to the Table 1 interpretation.

Circularity Check

3 steps flagged

Partial train–eval metric circularity: preference/aesthetic rewards are distilled from ImageReward, PickScore, HPSv2, and Aesthetic, then Table 1 re-grades on those same scorers.

specific steps
  1. fitted input called prediction [Appendix A.4 Eqs. (22)–(24); A.5 Eqs. (25)–(26); §3.4 Eqs. (14)–(16), (20); Table 1]
    "¯rp,pref = (˜rIR_p + ˜rPS_p + ˜rHPS_p)/3 ... For numerical assessment, we compute an aesthetic metric score ap for Ip and normalize it: ¯rp,aes = mm(ap; Dhigh) ... Fi = F_θold(x, ˆxi, Ii), ri = [ri_sem, ri_aes, ri_pref]. ... Ri = α Rideal_i + (1−α) Rcheb_i ... Method ... Aesthetic↑ PickScore↑ HPSv2↑ ImageReward↑ ... PRISM 0.299 6.318 22.526 0.305 1.285"

    Preference and aesthetic supervision targets are defined from ImageReward, PickScore, HPSv2, and an aesthetic scorer; the VLM judge is SFT-calibrated to those discretized scores and then used as the GSPO reward. Table 1 evaluates the optimized policy on Aesthetic, PickScore, HPSv2, and ImageReward—the same family the policy was trained to raise. Reported lifts on those four columns are therefore statistically encouraged by construction of the objective (metric-fitting), not an independent external test of preference/aesthetics. This does not void TIFA/CompBench, but it weakens the holistic Table-1 claim.

  2. fitted input called prediction [Appendix A.1 Eq. (22); A.2–A.3 data selection; Table 1 CLIPScore column]
    "We then define a unified quality score: Si = 0.25 ãi + 0.25 p̃i + 0.50 c̃i. The larger CLIPScore weight emphasizes image-text semantic alignment. We use Equation (22) for ranking samples in both Type-I and Type-II data construction. ... retain the top N1 = 10,000 pairs ... retain the top N2 = 5,000 high-quality pairs."

    Training pairs are ranked and filtered with a pool score that puts 50% weight on CLIPScore (plus PickScore and Aesthetic). The learned rewriting policy is thus selected/shaped on data already enriched for high CLIP/Pick/Aesthetic, then Table 1 again reports CLIPScore (and the same preference/aesthetic family). CLIPScore gains are partly a consequence of how the supervised corpus was constructed, not a pure out-of-distribution prediction.

  3. other [§3.4 Eqs. (12)–(14); Table 3 (Self-FB / PRISM vs Ext-VLM)]
    "A frozen copy of the initialized VLM is then used in feedback assessment mode to produce dimension-specific scores: Fi = F_θold(x, ˆxi, Ii), ri = [ri_sem, ri_aes, ri_pref]. ... Ext-VLM Ext RL 22.411 90.0 ... PRISM Self RL 22.526 91.4"

    Self-rewarding uses a frozen twin of the same SFT VLM as judge—classic self-evaluation structure. This alone is not full circularity: images still come from a frozen external T2I model, and Table 3’s Ext-VLM control reaches nearly the same PickScore/TIFA, so the twin-judge is not load-bearing for the entire gain. It does, however, couple policy and reward representation and compounds the metric-overlap issue above when the twin was distilled from the eval scorers.

full rationale

PRISM’s derivation is not tautological by definition: an external frozen T2I generator produces images, semantic feedback uses DSG-style VQA rather than the Table-1 scorers, and TIFA/T2I-CompBench are held-out compositional metrics that are not literally the Stage-II reward. Table 3 further shows a self-judge vs external-VLM-judge gap that is small, so pure self-reward twinning is not the main issue. The load-bearing circularity is narrower and checkable: Appendix A.4 builds the preference target as the average of normalized ImageReward, PickScore, and HPSv2, and aesthetic targets from an external aesthetic metric; Dscore/SFT distill the VLM judge to those scores; Stage II (Eqs. 14–20) optimizes the prompt policy against that judge’s hybrid reward; Table 1 then reports gains on Aesthetic, PickScore, HPSv2, and ImageReward—four of five columns that are constituents of the training reward. Those holistic gains are therefore partly metric-fitting (fitted objectives re-read as evaluation), not fully independent evidence of “holistic image quality and human preference.” Independent content remains on TIFA/CompBench and in the structured-diagnosis rewriting story, so this is partial circularity (≈5), not a full by-construction collapse.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on standard ML training assumptions, a frozen T2I generator, proxy quality scorers treated as supervision, and several hand-chosen multi-task and multi-objective weights. No new physical entities are postulated; the main invented constructs are the PRISM loop and the hybrid reward aggregator.

free parameters (5)
  • multi-task SFT loss weights λ_rew:λ_fb:λ_score:λ_fgr = 1:1:1:2
    Set by hand to 1:1:1:2 to emphasize feedback-guided rewriting; directly shapes the initialized policy.
  • hybrid reward dimension weights w and mix α = w=[0.5,0.25,0.25]; α in [0,1] (exact train α not uniquely fixed in main text)
    w=[0.5,0.25,0.25] emphasizes semantics; α mixes ideal-point vs Chebyshev terms. Chosen design knobs that define the scalar reward optimized by GSPO.
  • GSPO/sampling hyperparameters (K, T, top-p, ε_c, β) = K=8, T=0.9, top-p=0.9, ε_c=0.2, β=0.3
    Candidate count, temperature, clipping, and KL strength control exploration vs collapse in self-rewarding updates.
  • unified quality score weights for data filtering (0.25 Aesthetic + 0.25 PickScore + 0.50 CLIPScore) = 0.25/0.25/0.50
    Hand-set ranking formula that selects BeautifulPrompt and community pairs used as supervision (Eq. 22).
  • dataset size cutoffs N1, N2 and simplification schedule = N1=10000; N2=5000; simplification mix 30/25/30/15%
    Top-10k BeautifulPrompt pairs and top-5k community pairs plus stochastic degradation rates determine the training distribution.
axioms (5)
  • domain assumption Frozen off-the-shelf T2I generators (SD3.5 medium, FLUX.2-klein) are adequate environments for learning a transferable prompt policy.
    Problem formulation Eq. 1 and experimental setup treat G as fixed; policy quality is defined relative to these generators.
  • domain assumption Proxy scorers (CLIPScore/aesthetic metrics, ImageReward, PickScore, HPSv2) plus VLM QA over DSG-style atoms are valid stand-ins for semantic, aesthetic, and human preference quality.
    Appendix A.4–A.5 builds all numerical supervision and much of the evaluation stack from these proxies.
  • ad hoc to paper A frozen copy of the SFT VLM yields usable scalar rewards for on-policy improvement without an external preference model.
    Self-rewarding stage (Eqs. 12–20) depends on this judge; it is a methodological choice specific to PRISM’s loop.
  • domain assumption Group-normalized sequence-level advantages (GSPO) with KL to the SFT reference stabilize prompt-policy RL.
    Optimization follows Zheng et al. 2025 GSPO; standard modern RLHF-style assumption imported into prompt rewriting.
  • domain assumption Evaluation against the original user prompt (not the optimized prompt) measures preservation of user intent.
    Stated in §4.2; required to interpret metric gains as intent-preserving improvement rather than prompt drift.
invented entities (2)
  • PRISM unified prompt-optimizer / visual-judge VLM policy no independent evidence
    purpose: Single model that both rewrites prompts and emits structured multi-axis visual feedback used as reward.
    Core system object introduced by the paper; evidence is internal empirical performance, not an external physical discovery.
  • Hybrid ideal-point + Chebyshev visual-feedback reward Φ(r) no independent evidence
    purpose: Aggregate semantic/aesthetic/preference scores into one scalar that balances overall improvement and worst-axis penalty.
    Defined in Eq. 16; ablation Table 4 supports usefulness inside this paper’s setup only.

pith-pipeline@v1.2.0-grok45-kimik3 · 24525 in / 3864 out tokens · 78545 ms · 2026-07-31T17:21:33.899020+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of PRISM: Prompt Refinement via Image-grounded Self-rewarding Mechanism for Text-to-Image Generation." pith.science (2026). https://pith.science/paper/GSHM7C4I

@misc{pith2026260724353,
  author       = {Pith},
  title        = {Pith review of: PRISM: Prompt Refinement via Image-grounded Self-rewarding Mechanism for Text-to-Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSHM7C4I}},
  note         = {Machine review of arXiv:2607.24353}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-to-image generation models can synthesize high-quality images from natural language descriptions, but their performance remains highly sensitive to prompt formulation. Existing prompt optimization methods mainly rely on text-side rewriting, prompt expansion, or external reward signals, offering limited image-grounded diagnosis and weak support for learning reusable optimisation policies. In this paper, we propose PRISM, a Prompt Refinement framework via Image-grounded Self-rewarding Mechanism. PRISM closes the prompt-image-feedback loop by interpreting generated images with structured visual diagnosis and scoring them along semantic consistency, aesthetic quality, and human preference alignment. It first initializes a unified VLM through multi-task supervised fine-tuning, and then improves the prompt policy via self-rewarding optimization with a hybrid ideal-point and Chebyshev reward. Extensive experiments show that PRISM improves holistic image quality and fine-grained semantic alignment, while providing interpretable feedback for targeted prompt refinement. The code is available at https://anonymous.4open.science/r/PRISM-FF81.

Figures

Figures reproduced from arXiv: 2607.24353 by Guo Tang, Hao Wang, Hongjie Luo, Tianxu Wang, Ying Zhang.

Figure 1
Figure 1. Figure 1: Motivation of the proposed PRISM frame￾work. and reinforcement learning-based prompt adapta￾tion (Cao et al., 2023; Pavlichenko and Ustalov, 2023; Yang et al., 2025). While effective, most of them rely on textual patterns or scalar rewards, providing limited diagnostic guidance on why a generated image succeeds or fails. Recent visual￾feedback-based methods incorporate generated im￾ages into prompt refinem… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed PRISM framework. Stage I performs multi-task supervised initialization to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Iterative feedback refinement results across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples of PRISM. Each example shows the original prompt, initial generated image, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples of PRISM. Each example shows the original prompt, initial generated image, [PITH_FULL_IMAGE:figures/full_fig_p018_5.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

45 extracted references · 6 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2112.10741 , year=

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models , author=. arXiv preprint arXiv:2112.10741 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  3. [3]

    International conference on machine learning , pages=

    Generative adversarial text to image synthesis , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    Advances in neural information processing systems , volume=

    Photorealistic text-to-image diffusion models with deep language understanding , author=. Advances in neural information processing systems , volume=

  6. [6]

    International conference on machine learning , pages=

    Zero-shot text-to-image generation , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  7. [7]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Attngan: Fine-grained text to image generation with attentional generative adversarial networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  8. [8]

    International Conference on Learning Representations , volume=

    Sdxl: Improving latent diffusion models for high-resolution image synthesis , author=. International Conference on Learning Representations , volume=

  9. [9]

    International Conference on Learning Representations , year=

    PixArt- : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis , author=. International Conference on Learning Representations , year=

  10. [10]

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

    Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  11. [11]

    Advances in Neural Information Processing Systems , volume=

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation , author=. Advances in Neural Information Processing Systems , volume=

  12. [12]

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

    Test-time prompt refinement for text-to-image models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  13. [13]

    The Fourteenth International Conference on Learning Representations , year=

    VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis , author=. The Fourteenth International Conference on Learning Representations , year=

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Optimizing prompts for text-to-image generation , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

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

    Learning multi-dimensional human preference for text-to-image generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [17]

    2024 , howpublished =

    Stable Diffusion 3.5 Medium , author =. 2024 , howpublished =

    2024

  18. [18]

    International conference on learning representations , volume=

    Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation , author=. International conference on learning representations , volume=

  19. [19]

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

    Promptsculptor: Multi-agent based text-to-image prompt optimization , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages=

    2025

  20. [20]

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

    T2i-copilot: A training-free multi-agent text-to-image system for enhanced prompt interpretation and interactive generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  21. [21]

    arXiv preprint arXiv:2403.17804 , year=

    Improving text-to-image consistency via automatic prompt optimization , author=. arXiv preprint arXiv:2403.17804 , year=

    Pith/arXiv arXiv

  22. [22]

    IEEE Transactions on Visualization and Computer Graphics , volume=

    Promptmagician: Interactive prompt engineering for text-to-image creation , author=. IEEE Transactions on Visualization and Computer Graphics , volume=. 2023 , publisher=

    2023

  23. [23]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Best prompts for text-to-image models and how to find them , author=. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  24. [24]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    2023

  25. [25]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Self-rewarding large vision-language models for optimizing prompts in text-to-image generation , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  26. [26]

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

    Prompting hard or hardly prompting: Prompt inversion for text-to-image diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  27. [27]

    International conference on learning representations , volume=

    Pixart- : Fast training of diffusion transformer for photorealistic text-to-image synthesis , author=. International conference on learning representations , volume=

  28. [28]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  29. [29]

    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=

  30. [30]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  31. [31]

    Advances in neural information processing systems , volume=

    Cogview: Mastering text-to-image generation via transformers , author=. Advances in neural information processing systems , volume=

  32. [32]

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

    Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks , author=. Proceedings of the IEEE international conference on computer vision , pages=

  33. [33]

    arXiv preprint arXiv:2306.09341 , year=

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis , author=. arXiv preprint arXiv:2306.09341 , year=

    Pith/arXiv arXiv

  34. [34]

    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=

  35. [35]

    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=

  36. [36]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    Clipscore: A reference-free evaluation metric for image captioning , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  37. [37]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2507.18071 , year=

    Group sequence policy optimization , author=. arXiv preprint arXiv:2507.18071 , year=

    Pith/arXiv arXiv

  39. [39]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  40. [40]

    Publications Manual , year = "1983", publisher =

    1983

  41. [41]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  42. [42]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  43. [43]

    Dan Gusfield , title =. 1997

    1997

  44. [44]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  45. [45]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =