acceptodds
Withdrawn from ICLR 2027

Don't Generate Just to Discard: Prompt-Side Aesthetic Forecasting for Text-to-Image

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Achieving satisfactory text-to-image (T2I) results often requires costly trial and error: users repeatedly revise prompts, generate images, and evaluate them manually or with quality assessment models, wasting computation and waiting time on unsatisfactory outputs. This leaves a practical but underexplored question: can we assess a prompt's quality potential before generating any images? Existing prompt-side studies find meaningful correlations between prompt embeddings and image quality averaged across repeated generations, suggesting that expected generation quality can be partially forecast from text alone. Yet limitations remain at three levels. At the quantification level, they collapse stochastic outcomes into an average score, obscuring variation across images generated from the same prompt. At the modeling level, coarse prompt-to-score mappings do not capture contextual interactions or address the cost of obtaining reliable multi-sample supervision. At the interpretability level, average attribution makes it difficult to identify which prompt components improve or impair image quality. Motivated by these limitations, we formulate Prompt-Side Aesthetics Forecasting (PSAF), a pre-generation task that estimates a given generator's expected aesthetic response from prompt text alone, and instantiate it with AesPrompt. AesPrompt uses learnable Aesthetic Response Queries to distill distributional aesthetic evidence from multiple generated samples into prompt-side representations, together with a Chunk-Aware Prompt Adapter that captures contextual interactions in long prompts. We further contribute PSAF-540K, a dataset of 540K samples designed for the PSAF task. During dataset construction, adaptive Bayesian response sampling directs the limited generation budget toward prompts expected to provide the most informative supervision. With only 95.7M trainable parameters, approximately 0.3% of a 30B-scale generator, AesPrompt achieves a Spearman rank correlation (SRCC) of 0.951 and a pairwise ranking accuracy (PairAcc) of 0.944 without sampling images at inference time. In typical prompt exploration workflows, filtering with AesPrompt can reduce the number of rendered candidates and GPU time by 60%–85% while retaining high-quality prompts for final generation. The code, datasets, and demo are included in the supplementary materials.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Text-to-Image Generation, Prompt-Level Aesthetic Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_yXv6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces Prompt-Side Aesthetic Forecasting (PSAF), which predicts how visually appealing AI-generated images will be from their text prompts alone. The authors build a large dataset and develop AesPrompt, a model that estimates both expected image quality and how much it may vary. AesPrompt outperforms existing methods and helps select better prompts before generating images. However, the dataset distribution, scorer, and problem formulation have limitations.

Critical strengths
  • the proposed method can predict aesthetic quality before image generation, thus reducing the cost of rendering and discarding images. It requires only prompt text, without access to the generator’s internals, so adding little overhead

  • as getting reliable prompt-level labels requires generating many images per prompt, which is expensive. The paper's adaptive Bayesian response sampling optimizes the overall budget and quality and effectively cuts the cost for building the dataset

  • the supplmentary code seems to be comprehensive to reproduce the reported test metrics

Critical weaknesses
  • as the main data is from a single closed sourced image model (MJ), I have a major concern over the soundness of this work. Different image models, versions, are known to have different characteristics, preferences. The experimental data built on top of single source, single model doesn't prove it generalizes to other generators. In cross-generator adapation table 2 results, it shows the transfer is weak and it's worse on newer models. a generator-specific approach which requires per-generator training makes it less attractive in practice given how frequently new models are released and how slow/expensive it could be to collect generator specific data and retrain

  • Same applies to the HPS which seems to be the main scorer. Table4's results on pickscore/clip score which are both a few years old and the it's evaluating on stable diffusion 1.5 which is certainly an outdated model. The paper would be more convincing if it also evaluated with current scorers and current generators in 6.2 exp to understand if the gain is consistent with new models/new & different scorers

  • this problem formulation does not cover image to image part and agentic workflows where a multi-turn image editing conversation is the prior. both have fast growing usage and presence in different mainstream applications, and both are more challenging/interesting

  • text to image has the culture aspect consideration and aesthetic preferences vary across cultures. how does this approach account for this diversity, and could learning from a particular generator and preference scorer reinforce or amplify cultural biases, favoring some visual traditions over others?

Questions for authors
  • Is the HPS (v3?) the main scorer for training dataset? This work may be learning to predict HPSv3's own biases rather than general human aesthetics. would need a better (small scale) exp similar to 6.2 to understand if the gain is consistent with newer models and other new scorers (or human preference)

  • how does it compare with an agentic prompt-optimization work like GPT image 2.5 thinking (chatgpt.com/images)? basically an LLM to optimize the prompt iteratively to achieve best aesthetics with better generability (observe, improve)

Reviewer_PxfzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces prompt-side aesthetic forecasting, which predicts from the prompt text alone how good the images of one fixed T2I generator will be. The goal is to let users skip rendering weak prompts but direclty predicting one average score like prior prompt-side work, it predicts several statistics of the score distribution over many renders. The authors build a 540K dataset with adaptive Bayesian sampling that spends more renders on the most informative prompts. Their model AesPrompt adds a small adapter to a frozen CLIP text encoder and learns from image scores only during training. The authors show that AesPrompt clearly beats adapted text and multimodal backbones on their benchmark. The training-only image stage is the main reason for this gain.

In conclusion, this paper mainly shows:

  • Prompt quality can be forecast from text as a score profile or statistics instead of one mean.
  • Image evidence used only in training makes a text-only predictor much stronger.
  • Adaptive render allocation gives labels of similar quality with about half the renders.
Critical strengths
  • The paper sorts prior work clearly by what each method needs from the generator. Its training-only image stage is a sensible design. Removing this stage drops SRCC from 0.951 to 0.827.
  • Data collection under a render budget is studied with care. Adaptive allocation matches the label quality of eight renders per prompt with about four (Fig. 4 and appendix Table 8). The appendix also reports the accuracy and the calibration of each predicted statistic.
Critical weaknesses
  1. The paper never explains why we need to judge a prompt before generation. Many current systems already rewrite the user prompt with a built-in LLM before rendering to avoid weak prompt result in bad images [1,2,3]. So the text the users writes is NOT what the generator sees. The paper also says prior prompt-side work only predicts one average score. But it never measures how much the images of one prompt vary and never compares with a mean only model (in AesPrompt's design). The cited CEMS already learns prompt-level quality variance. A high correlation on its own bench does not show that the prediction is useful. This is exactly the open question of the cited Kim & Mils (2026)
  2. The 60~85% saving in renders and GPU time appears only in the abstract with no experiment behind it. It also ignores that the cost of building the predictor is hundres of thousands of rendered and labeled images and about 1k more samples for every new generator. The only downstream test picks among rewrites from two old prompt optimizers made for SD 1.x series. The gain on 200 prompts is small and has no error bar provided (Usually error bar is no needed but for such small difference I think it is important, or p value). Also: general prompt optimizers [4,5,6] are not tried here as well.
  3. Based on the prediction files in the supp material, SRCC is 0.95 on the MidJourney prompts but only 0.43 on the 282 LongT2IBench prompts. Top-10 Recall on LongT2IBench is onlhy 0.29. The generator and the scorer are the same for both source. The supp split file shows that LongT2IBench prompts are even in the training set. Also the paper only reports the pooled numbers and hides this weak subset, its long-prompt analysis also pools both sources and concludes that long prompts only hurt top-candidate retrieval. So the claims on long prompts, cost saving and generalization do NOT hold once the prompt source changes.
  4. The predictor does not transfer well to other generators because its zero-shot SRCC drops to 0.29~0.46 and reaches at most 0.61 after adaptation. It is also tied to the one scorer that gives its training labels, which matters because aesthetic and preference scorers often disagree with humans and with each other [7,8,9] and some of them only reach about 0.3 or even near-zero correlation with each other on generated images [10], while this property of aesthetic scorer are known, this paper never tested transfer to another scorer. Even which scorer gives the labels is unclear, because the appendix calls them human scores while the supplementary code actually uses HPSv3 model scores, and based on the supplementary code the model also learns from nine automatic scorers through extra heads, where two of these scorers are baselines in Table 1 and the paper never mentions this.

References

  • [1] J. Betker et al. Improving Image Generation with Better Captions. OpenAI technical report, 2023
  • [2] Qwen Team. Qwen-Image-2.0 Technical Report. arXiv:2605.10730, 2026
  • [3] PromptEnhancer: A Simple Approach to Enhance Text-to-Image Models via Chain-of-Thought Prompt Rewriting. arXiv:2509.04545, 2025
  • [4] O. Mañas et al. Improving Text-to-Image Consistency via Automatic Prompt Optimization. TMLR 2024
  • [5] S. Datta et al. Prompt Expansion for Adaptive Text-to-Image Generation. arXiv:2312.16720, 2023
  • [6] S.-Y. Yeh et al. TIPO: Text to Image with Text Presampling for Prompt Optimization. ICLR 2026
  • [7] Z. Chen et al. MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? NeurIPS 2025 Datasets and Benchmarks
  • [8] T. Lee et al. Holistic Evaluation of Text-to-Image Models. NeurIPS 2023 Datasets and Benchmarks
  • [9] Evaluating the Evaluators: Metrics for Compositional Text-to-Image Generation. arXiv:2509.21227, 2025
  • [10] S.-Y. Yeh et al. Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation. ICLR 2024
Questions for authors
  1. How many renders, annotations and GPU hours did PSAF-540K and each new-generator adaptation need? After how many forecast queries does the predictor pay back this cost?
  2. Are the main targets human scores or HPSv3 model scores? What role do the nine teacher-score heads in the supplementary code play in training? Does the predictor keep its ranking when another scorer labels the same images?
  3. In a system that rewrites prompts before rendering, would AesPrompt score the user prompt or the rewritten one? Does it still help when it selects among candidates from a built-in prompt enhancer?

Besides the questions above, please also refer to the weakness part.

Reviewer_xphJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes AesPrompt, a lightweight prompt-side aesthetic forecasting framework that predicts the potential quality of text-to-image (T2I) generation before image synthesis. By leveraging training-time cross-modal supervision, the method aims to reduce unnecessary generation costs while achieving strong prompt-ranking performance. The motivation is clear, and the experimental results demonstrate promising performance within the evaluated setting.

Critical strengths

The paper addresses a practical and well-motivated problem: reducing the computational and financial costs associated with repeated T2I generation. By evaluating prompts before generation, the proposed method can help filter low-quality candidates and improve generation efficiency. The method achieves strong quantitative results compared with existing baselines, demonstrating the potential of prompt-side aesthetic forecasting as an effective strategy for reducing unnecessary generation.

Critical weaknesses

A major concern is the generalization capability across different T2I generators. Commercial image-generation models may exhibit substantially different sensitivities to prompt wording, semantic structures, and stylistic descriptions. Consequently, training on data collected from a specific generator may introduce generator-specific biases and limit the model's ability to generalize to other generators. The reported cross-generator results also suggest that zero-shot transfer remains challenging. This limitation may restrict the practical applicability of the proposed method in real-world scenarios where users frequently switch between different generation services.

Questions for authors

How would AesPrompt perform when a new commercial T2I model becomes available? Given the generator-specific nature of prompt sensitivity, the proposed method may require additional adaptation data to achieve reliable predictions for the new model. However, collecting such data may require substantial API calls and image-generation costs, potentially undermining the paper's central motivation of reducing generation overhead. Could the authors clarify how the method can be efficiently adapted to previously unseen generators, and whether the resulting data-collection and adaptation costs are justified by the subsequent savings in generation costs?

Reviewer_jmb5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies prompt-side aesthetic forecasting for a fixed text-to-image generator. AesPrompt distills image-derived evidence into a text-only predictor using aesthetic response pretraining, learnable response queries, and a chunk-aware adapter. PSAF-540K combines prompt-level targets with long-prompt, validity, and phrase-contribution annotations. The main benchmark reports substantial gains over adapted prompt-only baselines, with additional experiments on cross-generator adaptation, phrase attribution, and rendering-budget allocation.

Critical strengths
  1. A useful pre-generation decision problem with a substantial reported improvement. In Table 1 (p. 8), SRCC improves from 0.885 for the FLUX text stack to 0.951, and Top-10 Recall improves from 0.554 to 0.695. This directly addresses whether rendering can be concentrated on promising prompts.
  2. Evidence that the proposed training signals matter. Removing ARP reduces SRCC to 0.827, while removing the ranking loss reduces it to 0.814 (Table 3, p. 9). The 32-render reference experiment also provides a concrete way to evaluate sampling efficiency rather than relying only on predictor accuracy.
Critical weaknesses
  1. The benchmark does not yet establish independent aesthetic validity or a fully auditable split. Figure 2 (p. 5) states that the main targets are MidJourney-derived and HPSv3-supervised. Agreement with these targets demonstrates prediction of that evaluator, but is not sufficient evidence of general human aesthetic preference. The paper reports a fixed 80/10/10 prompt split (p. 7), but does not establish in the supplied PDF that original prompts, long variants, and phrase-derived records are grouped before splitting. This is a leakage risk requiring verification, not evidence that leakage occurred. The claimed 60–85% GPU savings also lack a traceable end-to-end timing protocol here. These gaps make the unusually high headline accuracy difficult to translate into practical value.
  2. Distributional targets and efficiency comparisons require better specification. Equation 1 uses empirical maxima and minima as bounds on a latent response, although extrema depend on the number of samples and do not bound future draws. Adaptive allocation changes that sample count across prompts. The paper therefore needs to distinguish response variation from sample-budget effects. Moreover, 95.7M is explicitly the trainable parameter count, while a frozen OpenCLIP text tower remains at inference (p. 6); Table 1 needs a consistent distinction between total and trainable parameters. The missing referenced appendices prevent checking the acquisition rule, distribution calibration, and adapted-baseline training details.
Questions for authors
  1. Can you provide the referenced appendices A.1–A.5 and a group-level split audit showing that all derivatives of each source prompt remain in one partition? Does the main improvement persist under a semantic-family holdout?
  2. How do you control sample-count effects in maximum/minimum prediction under adaptive sampling? Please report fixed-count or quantile-based evaluation and per-statistic calibration on independently rendered held-out prompts.
  3. Can you provide an independent human-preference test and end-to-end rendering-budget curves, including predictor and adaptation costs? Please separate total inference parameters from trainable parameters for every baseline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.