Pith. sign in

REVIEW 3 major objections 4 minor 16 references

The paper claims that each failed part of a text-to-image prompt is best fixed by its own repair language: type-specific prompts, compiled into one rewrite and gated by accept-or-revert, beat a single uniform expansion in every benchmark–ge

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 14:33 UTC pith:4V5M5WOI

load-bearing objection Type-aware repair allocation is a real advance; the headline gap likely shrinks under independent judges, but the core result holds. the 3 major comments →

arxiv 2607.18724 v1 pith:4V5M5WOI submitted 2026-07-21 cs.AI

One Rewrite to Fix Them All? Type-Aware Repair Allocation for Text-to-Image Prompt Optimization

Haoyue Liu , Xiaoyu Ma , Ye Chen , Shuguang Cui , Xiaoying Tang This is my paper
classification cs.AI
keywords text-to-image generationprompt optimizationsemantic faithfulnessatomic repair allocationtype-conditioned repairvision-language feedbackaccept-or-revert gateprompt rewriting
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 generators fail prompts in different ways — wrong counts, swapped attributes, tangled relations, illegible text — and this paper tries to establish that each failure calls for its own repair language: a prompt optimizer should route every diagnosed failure to a type-specific fix rather than absorb everything into one uniform expansion. The proposed framework, TARA, decomposes the prompt into atomic checkable propositions, scores each with a vision-language judge, assigns each failure one of eight repair operators (word-and-digit counts with visible separation, quoted strings with legible typography, left/right/above/below anchors for relations), compiles the local repairs into one fused prompt, and adopts the single regenerated image only if its semantic score improves, otherwise reverting to the original. Formally, uniform expansion is shown to be the degenerate single-operator special case of this atomic repair allocation (Eq. 3), turning the comparison with prior optimizers into a controlled test of type-awareness. Across the DSG and TIFA benchmarks and four frozen generators, TARA reports the best semantic accuracy in all eight cells — +5.6 and +2.6 points over the strongest visual-feedback baseline — while preserving image quality and running fastest, with the ranking surviving four held-out re-scorers and a 100-case human study. If the paper is right, the bottleneck in prompt optimization is not how much you rewrite but whether the rewrite matches the failure mode, and one carefully gated regeneration can be enough.

Core claim

Atomic visual feedback should decide how each failure is repaired, not merely that a rewrite happens. TARA decomposes each prompt into atomic propositions, scores them with a frozen vision-language judge (Eq. 1), routes failures through a fixed category-to-type table to one of eight repair operators, and compiles the local repairs into one fused prompt (Eq. 2). Eq. (3) states the structural claim: uniform expansion is the type-invariant special case of this allocation, making prior optimizers special cases of it. Empirically, TARA leads all eight benchmark–generator cells, and the ranking survives four held-out re-scorers and a 100-case human study.

What carries the argument

Type-conditioned repair with routing is the engine. A fixed category-to-type table maps each failed proposition to one of eight repairable types — missing object, count, attribute, relation, action, text, style, global — and each type has its own repair dialect: word-and-digit counts with separation, exact quoted text, spatial anchors, pose cues. The allocator routes each failure; the compiler fuses a text-only portfolio (uniform, typed, minimal-append, re-seed if near-correct) into one prompt. An accept-or-revert gate adopts the single regenerated image iff its score improves, else reverts. Eq. (3) makes uniform expansion the type-invariant special case, so the comparison to prior optimizer

Load-bearing premise

Everything rests on one assumption: the AI judge that checks whether an image satisfies the prompt is accurate, because the same model family both diagnoses failures, decides whether to accept the repaired image, and defines the headline score (Appendix A.1 notes judge and rewriter share that family) — if the judge is biased toward TARA's rewording, the routing decisions and the reported +5.6/+2.6 gains could be artifacts of the judge rather than real semantic improvement.

What would settle it

Take the saved images from TARA and VisualPrompter on a random subset of prompts and have human annotators answer the same atomic yes/no questions the judge answered, then recompute semantic accuracy from human labels alone; if the all-eight-cells lead, or the +5.6/+2.6 gaps, shrinks to noise, the central claim collapses. A cheaper intermediate check: swap the judge inside the accept-or-revert gate for a different model family and test whether the repairs it accepts still genuinely improve under that independent judge.

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

If this is right

  • Typed routing is worth real points: with the image budget held fixed, dropping the explicit typed-expansion branch alone costs 3.8 points of semantic accuracy on DSG, and the full method leads all eight benchmark–generator cells.
  • One gated regeneration beats unconditional rewriting: at most two image generations per prompt, with the accept-or-revert gate ensuring a failed repair reverts to the original rather than committing a regression.
  • Repair language transfers across generators: the same fixed templates lift accuracy on four frozen open generators and on two closed commercial systems shown in the appendix, without training or white-box access.
  • Short typed rewrites win on cost too: roughly 18-word fused prompts reach higher accuracy than roughly 30-word uniform expansions while cutting end-to-end time from 20.0s to 16.0s per prompt.
  • Gains are broad, not traded: TARA improves over the raw prompt on all eight semantic categories and matches the strongest baseline on seven of eight, so type-aware repair does not fix one failure mode at another's expense.

Where Pith is reading between the lines

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

  • Editorial extension: since the gate adopts only when the judge's score rises, TARA's ceiling is the perceptual fidelity of that judge — and since the judge shares a model family with the rewriter, the strongest follow-up would run the gate itself (not just the paper's post-hoc re-scorers) on an independent judge family or human labels.
  • The hand-built repair templates read as codified knowledge of how diffusion models fail — word-and-digit counts, separated instances, quoted strings; learning these repair dialects from data is a testable extension that would let routing adapt to a new generator's systematic failure modes.
  • The paper fixes one global threshold for deciding when a near-correct image should be re-seeded, yet its own per-generator tables show the same failure type is far more repairable on stronger generators — per-generator or per-type thresholds are a plausible cheap improvement left unexplored.
  • TARA's four-stage decomposition separates diagnosis from repair language, suggesting a modular design in which each stage — judge, allocator, compiler, gate — can be upgraded independently without retraining the rest.

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 / 4 minor

Summary. This paper formulates text-to-image prompt optimization as atomic repair allocation: diagnose failed atomic propositions, allocate each failure to a type-conditioned repair operator, compile the local repairs into one prompt, and adopt the repaired image only if a VLM judge reports an improved semantic score. The TARA framework is compared with five existing optimizers on DSG-1k and TIFA v1.0 across four generators (SD 1.5, SD 2.1, Flux-dev, Janus-Pro). The authors report that TARA achieves the best average semantic accuracy in all eight benchmark–generator cells, with +5.6/+2.6 points over VisualPrompter on DSG/TIFA, while preserving CLIP/aesthetic scores and running at 16.0 s/prompt versus 20.0 s. The formulation is clear, and Eq. (3) usefully casts uniform expansion as a type-invariant special case. The empirical apparatus is unusually thorough for this subfield: three seeds, matched image budgets, prompt-clustered bootstrap intervals, four held-out re-scoring judges, a full ablation, and a 100-case human study. The paper is also candid about residual failures (Appendix B.4) and about the fact that the judge and rewriter share a model family (Appendix A.1).

Significance. If the reported gains hold, TARA is a practical contribution: it improves semantic faithfulness without training, generator access, or extra image generations, and its separation of diagnosis, allocation, compilation, and adoption is conceptually clean. The fixed repair templates, the matched image budget, and the explicit runtime accounting are strengths, as is the fact that the main comparison is not seed luck (three seeds, prompt-clustered bootstrap CIs excluding zero). The principal concern is that the headline comparison is intertwined with the accept-or-revert gate selecting on the same VLM judge that defines the evaluation metric; the held-out re-scorers and human study reduce but do not eliminate this coupling. Because this coupling bears directly on the central claim, I would need a revision before accepting.

major comments (3)
  1. [§3.1, §2.3, Algorithm 1 line 12] The central comparative metric is the same function the method optimizes. Eq. (1) defines semantic accuracy s(x) with a VLM judge, and Algorithm 1 line 12 adopts the repaired image iff s(hat_x) > s(x0) using that same judge. The reported score for TARA is therefore max(s(x0), s(hat_x)) over two noisy measurements of the same judge. Even a useless rewrite can produce a positive expected gain from judge noise alone. The paper's own Table 1 is consistent with this: re-scoring saved images with four held-out VLMs lowers the mean overall gain from +4.11 to +2.4–+3.2. I do not see this as disqualifying, but the claim that the +5.6/+2.6 advantage over VisualPrompter is caused by type-aware repair needs a control. Please add a VisualPrompter-plus-gate variant that applies the same accept-or-revert rule to VisualPrompter's single rewrite under the matched budget, and/or report a variant in which
  2. [§3.4, Figure 5, Appendix A.1] The human study does not break the judge coupling for the headline comparison. It compares TARA against the original prompt only, not against VisualPrompter, and it asks for preference rather than scoring the atomic semantic-accuracy questions. The four re-scorers in Table 1 are all VLM judges from overlapping or related model families (three Qwen variants and Gemma), so they may share systematic biases to an unknown degree. An independent assessment would require either human atomic-question verification on a sample of both TARA and VisualPrompter images, or at least a human preference test between TARA and VisualPrompter. As it stands, the 100-case study cannot establish that the +5.6/+2.6 gap over VisualPrompter survives an independent judge.
  3. [§3.2, Tables 2–3] The headline 'best in all eight benchmark–generator cells' is supported only at the aggregate level. The prompt-clustered bootstrap CIs are reported for the DSG and TIFA averages, not per generator cell. Within cells there are negative TARA−VisualPrompter entries in Tables 2 and 3 (e.g., Flux-dev/DiffusionDB −1.3, Janus-Pro/TIFA PartiPrompt −0.2, PaintSkill −1.0). These may well be noise, but per-cell intervals or a per-cell significance statement are needed to substantiate the 'all eight cells' wording. If the claim is intentionally about average per benchmark, the text should say so explicitly.
minor comments (4)
  1. [Eq. (2), §2.2] The notation R(d0) is used before it is formally defined; the recovery set is described in §2.2, but a compact definition at first use would improve readability.
  2. [Appendix C.2, Table 15] The threshold sweep is useful, but the default row is not bitwise comparable to the other rows because image generation is not guaranteed to be deterministic across GPU sessions. This caveat should appear in the main text or in the table caption, not only in the appendix prose.
  3. [Figure 5] The human study reports only percentages for 100 cases. Reporting a confidence interval and, ideally, an inter-rater agreement statistic would make the result more interpretable.
  4. [Table 12] The prompt-length values for NeuroPrompts, Promptist, and BeautifulPrompt are identical across all four generator columns. Please confirm that these are generator-independent aggregates and clarify the caption, since the table header implies per-generator values.

Circularity Check

1 steps flagged

No derivation chain is circular, but the semantic repair gate selects on the same VLM judge that defines the headline metric, partially reducing the reported gain to selection on the evaluation metric.

specific steps
  1. other [§2.1 Eq. (1); §2.3 'semantic repair gate'; Algorithm 1 line 12]
    "Eq. (1): s(x) = 1/n |{i:V(x,q_i)=CORRECT}|. §2.3: 'after generating exactly one prescribed repair ˆx=G(ˆp), it adopts the repair iff s(ˆx)> s(x0) and otherwise reverts to (p, x0).'"

    Eq. (1) defines the reported semantic accuracy s(x) using judge V; §2.3/Algorithm 1 make the same s the accept/reject rule for the single repaired image. Hence TARA's reported output is definitionally argmax_{x in {x0, xhat}} s(x) under that same V. Because V is a noisy measurement, taking the higher of two noisy scores yields a positive expected gain even if the rewrite adds no real semantics, and this gain grows with judge noise. VisualPrompter has no comparable gate, so part of the +5.6/+2.6 headline (and the +3.6 gate-ablation effect) is selection on the evaluation metric rather than type-aware repair. Held-out re-scorers are correlated VLM judges and do not break the coupling; the human study compares against the original prompt, not against VisualPrompter under the atomic metric.

full rationale

This paper is an empirical systems paper, not a derivation, so most of the claimed derivation chain is non-circular: the formulation of atomic repair allocation is a re-description, uniform expansion as a special case is a definitional statement, and the benchmark comparisons use external DSG/TIFA datasets and external baselines. There is no load-bearing self-citation: VisualPrompter, DSG, and TIFA are independent prior works, and no uniqueness theorem or ansatz is imported from the authors' own earlier papers. The one genuinely circular element is the semantic repair gate: the metric being reported (Eq. 1) is the same function used to decide whether to accept the repaired image (Algorithm 1 line 12 / §2.3). This makes the reported semantic accuracy, by construction, the maximum of two noisy VLM-judge measurements, so a positive expected gain over a non-gated baseline is guaranteed even for a useless rewrite. The paper partially mitigates this with four held-out re-scorers and a 100-case human study, which provide independent evidence that the ranking is not solely an artifact of the main judge; the held-out re-scorers shrink the gap from +4.11 to +2.4–+3.2, and the human study compares TARA to the original prompt rather than to VisualPrompter under the same metric. The existence of these checks keeps the paper from being fully circular, but the headline gain still contains a component that is statistically forced by the gate. Score 4.0 reflects partial, not total, circularity: the central type-aware-repair claim has independent content and independent checks, but one of its headline 'predictions' is partially constructed by selecting on the evaluation metric.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The method adds no invented physical or formal entities: the eight error types and repair operators are bookkeeping/diagnostic categories, not new objects with independent falsifiable handles. The genuinely hand-tuned content is the repair-template language (Table 9) and the τ threshold; both are fixed before evaluation and disclosed, but they are the real 'parameters' of the method. The central claim is otherwise grounded in external benchmarks (DSG, TIFA) and external baselines.

free parameters (2)
  • Re-seed threshold τ = τ = 0.72, swept over [0.60, 0.84], fixed across generators and benchmarks
    Hand-chosen cutoff for admitting the re-seed candidate when s(x0) ≥ τ. Appendix C.2 sweeps the range and shows stable gains (+8 to +10 points), so it is a weak parameter, but still a hand-set constant.
  • LLM sampling temperature / top-p for rewrite candidates = temperature 0.7, top-p 0.8; scoring at temperature 0
    Chosen by hand for candidate diversity (Appendix A.1). Not fitted to data, but part of the method's configuration and fixed per request.
axioms (3)
  • domain assumption A VLM judge (Qwen3.5-9B, plus four held-out re-scorers) answers atomic yes/no propositions correctly, and the resulting semantic score s(x) faithfully measures prompt faithfulness.
    Eq. (1) defines the headline metric and the gate; the entire pipeline — diagnosis, routing, accept-or-revert, and evaluation — rests on judge fidelity. §2.1, §3.1, Appendix A.1.
  • domain assumption DSG-style dependency pruning correctly handles absent entities, and pruned propositions counting zero-in-numerator / stay-in-denominator is the right scoring protocol.
    §2.2 — the recovery step R(d0) and the comparable metric for all methods depend on the DSG dependency graph and its pruning semantics.
  • ad hoc to paper Every semantic failure can be typed into one of eight categories (entity, count, attribute, relation, action, text, style, global), and the hand-authored repair templates (Table 9) are effective across all four generators.
    §2.2, Table 9 — the fixed category-to-type mapping and repair-language templates are author-designed, not learned or fitted; Table 14 shows repairability varies strongly across generators, so template generality is assumed rather than demonstrated.

pith-pipeline@v1.3.0-alltime-deepseek · 24082 in / 21022 out tokens · 172254 ms · 2026-08-01T14:33:19.273339+00:00 · methodology

0 comments
read the original abstract

Text-to-image (T2I) generators often fail to follow their prompts faithfully, producing wrong counts, swapped attributes, ambiguous relations, and illegible text. Prompt optimization repairs such failures by rewriting the user prompt, requiring no generator retraining, and has yielded promising results. However, existing optimizers absorb heterogeneous failures into one uniform prompt expansion, even though each calls for different repair language. We formulate semantic prompt optimization as atomic repair allocation: each failed proposition is routed to a type-conditioned repair operator before the resulting local constraints are compiled into one executable prompt. We instantiate this formulation in the training-free Type-Aware Repair Allocation (TARA) framework, which separates diagnosis, allocation, compilation, and a semantic repair gate, an accept-or-revert controller over exactly one prescribed repair that prevents semantic regressions. Extensive experiments on DSG and TIFA across four frozen generators demonstrate that TARA achieves the best semantic accuracy in all eight benchmark-generator cells, improving over VisualPrompter by 5.6 and 2.6 points on DSG and TIFA, respectively, while maintaining image quality and running fastest in our matched local setting at 16.0 seconds versus 20.0 seconds per prompt.

Figures

Figures reproduced from arXiv: 2607.18724 by Haoyue Liu, Shuguang Cui, Xiaoying Tang, Xiaoyu Ma, Ye Chen.

Figure 1
Figure 1. Figure 1: Type-agnostic expansion (VisualPrompter, left) versus atomic repair allocation (TARA, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TARA. It diagnoses failed atomic propositions, allocates type-conditioned [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on SD v1.5. TARA (green) fixes the count (top) and the missing [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-type semantic accuracy of seven methods over eight categories, aggregated over 24 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human preferences on 100 sampled cases under VisualPrompter’s three-way protocol: TARA / Tie / Original. Following the human-evaluation protocol of Vi￾sualPrompter (Wu et al., 2026), we compare TARA-optimized outputs against their original￾prompt counterparts on 100 sampled cases from two representative generators, allowing raters to express equal preference. Human judgments corroborate the automatic resul… view at source ↗
Figure 6
Figure 6. Figure 6: Full-scale DSG ablation across four generators and three seeds; diamonds denote means [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Worked example of type-aware repair (SD 1.5, DrawText). The initial image is decom [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on SD 1.5 / DSG. Each row is one prompt; columns are the raw [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Original vs. TARA on Stable Diffusion v1.5 (DSG), grouped by source category. Left: [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Original vs. TARA on Stable Diffusion v2.1 (DSG). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Original vs. TARA on FLUX-dev (DSG). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Original vs. TARA on Janus-Pro (DSG). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Original vs. TARA on Stable Diffusion v1.5 (TIFA), grouped by source category. Con [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Original vs. TARA on Stable Diffusion v2.1 (TIFA). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Original vs. TARA on FLUX-dev (TIFA). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Original vs. TARA on Janus-Pro (TIFA). Conventions as in Figure 9. [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Adaptation to commercial online generators—top: Doubao (text “Support Skywriters” [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Failure cases (DSG). Left: original; right (red outline): TARA. Numbers are VLM [PITH_FULL_IMAGE:figures/full_fig_p030_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Distribution of all eight diagnosed error types over failed atomic propositions, aggregated [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Optimized-prompt length over all prompts (DSG and TIFA, four generators). TARA’s [PITH_FULL_IMAGE:figures/full_fig_p031_20.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

16 extracted references · 5 linked inside Pith

  1. [2]

    Janus-Pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811,

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-Pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811,

    Pith/arXiv arXiv

  2. [6]

    FRAP: Faithful and realistic text-to-image generation with adaptive prompt weighting.arXiv preprint arXiv:2408.11706,

    Liyao Jiang, Negar Hassanpour, Mohammad Salameh, Mohan Sai Singamsetti, Fengyu Sun, Wei Lu, and Di Niu. FRAP: Faithful and realistic text-to-image generation with adaptive prompt weighting.arXiv preprint arXiv:2408.11706,

    Pith/arXiv arXiv

  3. [8]

    Design guidelines for prompt engineering text-to-image generative models

    Vivian Liu and Lydia B Chilton. Design guidelines for prompt engineering text-to-image generative models. InProceedings of the 2022 CHI conference on human factors in computing systems, pp. 1–23,

    2022

  4. [9]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi

    URLhttps://mlanthology.org/tmlr/2024/manas2024tmlr-improving/. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. InProceedings of the 2023 Conference on Empirical Methods in Natural La...

    2024

  5. [10]

    Best prompts for text-to-image models and how to find them

    Nikita Pavlichenko and Dmitry Ustalov. Best prompts for text-to-image models and how to find them. InProceedings of the 46th International ACM SIGIR Conference on Research and Devel- opment in Information Retrieval, pp. 2067–2071,

    2067

  6. [11]

    A systematic survey of prompt engineering in large language models: Techniques and applica- tions.arXiv preprint arXiv:2402.07927, 1,

    Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applica- tions.arXiv preprint arXiv:2402.07927, 1,

    Pith/arXiv arXiv

  7. [13]

    PromptCharm: Text-to-image generation through multi-modal prompting and refinement

    Zhijie Wang, Yuheng Huang, Da Song, Lei Ma, and Tianyi Zhang. PromptCharm: Text-to-image generation through multi-modal prompting and refinement. InProceedings of the 2024 CHI conference on human factors in computing systems, pp. 1–21,

    2024

  8. [15]

    TIPO: Text to image with text presampling for prompt optimization.arXiv preprint arXiv:2411.08127,

    Shih-Ying Yeh, Yi Li, Sang-Hyun Park, Giyeong Oh, Xuehai Wang, Min Song, Youngjae Yu, and Shang-Hong Lai. TIPO: Text to image with text presampling for prompt optimization.arXiv preprint arXiv:2411.08127,

    arXiv

  9. [16]

    Repair time

    URLhttps: //icml.cc/virtual/2026/poster/62440. 13 CONTENTS OF THEPAPER 1 Introduction 1 2 TARA 3 2.1 Overview and Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2.2 Typed Visual Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2.3 Typed Repair Compilation and Output Adoption . . . . . . . . . . . . . ....

    2026

  10. [2014]

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy

    URLhttp://arxiv.org/abs/1312.6114. Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick- a-pic: An open dataset of user preferences for text-to-image generation. InAdvances in Neural Information Processing Systems 36, NeurIPS 2023,

    Pith/arXiv arXiv 2023

  11. [2017]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    11 Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    Pith/arXiv arXiv

  12. [2021]

    URL https://aclanthology.org/2021.tacl-1.47/

    doi: 10.1162/tacl a 00397. URL https://aclanthology.org/2021.tacl-1.47/. Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, volume 34, pp. 8780–8794,

    doi:10.1162/tacl 2021

  13. [2022]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInterna- tional Conference on Learning Representations, 2021a. URLhttps://openreview.net/ forum?id=St1giarCHLP. Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic different...

    2023

  14. [2023]

    Beautiful- Prompt: Towards automatic prompt engineering for text-to-image synthesis

    Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beautiful- Prompt: Towards automatic prompt engineering for text-to-image synthesis. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pp. 1–11,

    2023

  15. [2024]

    CLIPScore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. InProceedings of the 2021 conference on empirical methods in natural language processing, pp. 7514–7528,

    2021

  16. [2026]

    net/forum?id=hIwVFRLaFy

    URLhttps://openreview. net/forum?id=hIwVFRLaFy. Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. InProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 2096–2105,

    2096