A Picture is Worth a Thousand Prompts? Efficacy of Iterative Human-Driven Prompt Refinement in Image Regeneration Tasks

1 Introduction

The rise of AI-generated content on online platforms has made it crucial to investigate how this type of content is created, specifically through the iterative processes of image generation and regeneration. While prior work has explored AI-led iterative refinement, this paper highlights the human's leading role in refining prompts and improving outcomes through their own judgment and control. The field of generative artificial intelligence (GenAI) has recently seen significant advancements, particularly in the development of text-to-image (txt2img) models. These models provide an easy and fast process for creating high-quality artwork. Among the notable txt2img models contributing to this trend are Midjourney [MidJourney, 2024], DALL-E 3 [Betker et al., 2023], and Stable Diffusion 3 [Esser and others, 2024]. While these models enable a more accessible way to generate high-quality and visually appealing images, creating artwork with this method (specifically, image generation) is usually iterative. A user starts with a concept, formulates a prompt for the txt2img model, and uses the prompt as input to the

Extended paper: https://arxiv.org/abs/2504.20340

model to obtain the desired image. If the obtained image is unsatisfactory, the user repeatedly refines the prompt until the desired result is achieved or the user abandons the task.

Image regeneration through prompt refinement refers to a task where a user iteratively edits their prompt with the goal of recreating a visual based on some target image or visual style. This iterative process illustrates human-AI interaction techniques, where the user and AI collaborate to achieve optimal outputs. By iteratively refining prompts, users actively guide the creative process, demonstrating the potential of human-AI collaboration to bridge gaps between technical capabilities and artistic intent. This concept of image regeneration through iterative prompt refinement has numerous practical applications, such as bypassing prompt marketplaces, educating novice users, restoration of lost or damaged art pieces [Trinh et al., 2024; Tang et al., 2024; Oppenlaender et al., 2024; Kulkarni et al., 2023; Liang et al., 2023]. Despite these potential applications, limited research exists on how humans can improve image quality through iterative prompt refinement.

Iterative refinement processes have been utilized by humans in a wide domain of tasks such as writing, programming, and design. Flower [Flower, 1981] provides a model on how writers plan, create, and revise their work iteratively. Madaan [Madaan et al., 2024] show that iterative refinement is effective in significantly improving outputs in text and code generation tasks through self-feedback mechanisms. Møller and Aiello [Møller and Aiello, 2024] show that stepwise prompt refinement can show improvement in text summarization tasks. Du [Du et al., 2022] provides the R3 framework that has demonstrated the effectiveness of iterative revision in producing high-quality textual outputs by incorporating user feedback at each stage of the revision. For txt2img generation, automatic prompt optimization systems have demonstrated substantial improvements in image quality by refining prompts systematically [Ma˜nas et al., 2024]. Building on these findings, this paper explores how iterative refinement impacts human-guided prompt optimization in image regeneration tasks.

In image regeneration and comparison, image similarity metrics (ISMs) such as Perceptual Similarity [Zhang et al., 2018], CLIP scores [OpenAI, 2021; Wang et al., 2023], and ImageHash [Buchner, 2024] can provide objective feedback on the likeness between two images [Saharia et al., 2022;

Zhang et al., 2018]. However, these ISMs' alignment with humans' subjective judgment remains untested. Since humans are the ultimate decision-makers in creative workflows, it is critical to ensure that ISMs align with subjective human evaluations. Humans make the final call on whether AIgenerated outputs meet their intended purpose, making human agreement essential to validate the reliability and practical applicability of ISMs. To address this, we first seek to understand the alignment of ISMs with human perception by comparing the objective rankings provided by these metrics to users' subjective rankings of the similarity between generated images and target images. This evaluation is vital for determining the feasibility of using ISMs as reliable feedback tools in iterative prompt refinement workflows.

Previous work by Trinh [Trinh et al., 2024] has studied how humans' inference may compare to machine inference (i.e CLIP interrogator), and shows that while humans are able to infer prompts and generate similar images, their efforts were not as effective as using the original target prompt. However, this previous research was limited to single-shot inference, where participants had only one attempt to generate a prompt. Furthermore, image similarity metrics (ISMs) were employed to define a threshold for successful inference, considering the task complete if the generated image achieved an ISM score above the threshold. In contrast to this prior work, our study seeks to examine how participants improve image regeneration performance when allowed multiple iterations to refine prompts. Instead of using ISMs solely to determine task completion, we leverage these metrics to quantify the iterative improvements in image regeneration. This allows us to investigate the effectiveness of iterative refinement as an approach to improving human-guided AI image generation.

From these research gaps and motivations, we conducted an experiment with human subject participants to assess their improvement in an image regeneration task This study also serves to provide additional insights on the alignment of different ISMs with subjective human assessment. Our contributions in this paper are as follows:

1. Survey Deployment: We conducted an in-person experiment with 20 human subject participants from our host institution.
2. SurveyData: Each participant conducted iterative prompt refinements for 10 target images over 10 iterations each, generating a total of 2000 prompts.
3. Data Evaluation: We utilized a comprehensive set of metrics in our evaluation, including Intraclass Correlation Coefficient to gauge the alignment of ISMs to human assessment. A mixed-effects model was used to analyze ISMs in quantifying the performance improvement of the iterative prompt refinement task. Additionally, we assess the iteration at which the highest user-ranked images were generated, as an additional metric to quantify improvement in iterative prompt refinement.

2 Background &RelatedWork

2.1 AI Image Generation

Modern txt2img systems often pair a language encoder (e.g., a Transformer-based or large language model) with a generative model, such as a diffusion model, a Generative Adversarial Network (GAN), or a Variational Autoencoder (VAE) [Jia et al., 2024]. In diffusion-based approaches (e.g., Stable Diffusion [Rombach et al., 2022]), the model iteratively denoises a latent representation conditioned on the text embedding, ultimately decoding it into pixel space. GANbased frameworks, by contrast, feature a generator trained adversarially against a discriminator [Goodfellow et al., 2020], driving the production of increasingly realistic images. VAEs encode input data into a latent space and then decode the latent vectors back into images, facilitating synthesis from sampled noise [Jia et al., 2024].

Despite these technical advances, small prompt modifications can produce markedly different outputs [Trinh et al., 2024]. Figure 1 demonstrates how substituting or reordering modifiers (e.g., bright colors, oil painting) in a prompt describing a dog can lead to substantial stylistic variations. This sensitivity underscores the need for systematic techniques, such as iterative prompt refinement, to achieve precise user objectives.

2.2 Iterative Prompt Refinement

While image generation often involves exploratory creativity, image regeneration introduces a more structured approach: recreating a specific target visual through iterative prompt refinement. This task requires users to refine their prompt iteratively, guided by feedback, either in the form of subjective assessment by the user or objective values such as ISMs, to achieve closer alignment with the target visual.

As illustrated in Figure 2, the process begins with the user analyzing the target image (Step 1) to identify key visual elements and features that need to be replicated. Based on this analysis, the user creates a text prompt (Step 2) and submits it to the txt2img generation model (Step 3), which produces an initial image. Following each generation, the process involves a subjective similarity assessment (Step 4a), where the user evaluates how closely the generated image matches the target image and refines the prompt accordingly. Additionally, a potential objective similarity assessment (Step 4b) uses image ISMs to provide a quantified feedback score. Steps 2, 3, 4a, and 4b are repeated iteratively, while the user edits the prompt in each iteration, and by the 10th iteration, the goal is to generate an image that closely aligns with the target image, reflecting improvements guided by human judgment and additional ISM feedback.

Image regeneration through prompt creation by humans can have several use cases. Trinh [Trinh et al., 2024] presented human prompt inference as a way to bypass prompt marketplaces, in turn questioning the validity of these marketplaces as business models. Image regeneration through iterative prompt refinement is an alternative for users to recreate target images without relying on purchased prompts. Additionally, iterative prompt refinement can also enable nonexpert users, such as hobbyists or novice designers, to engage and supplement their creative abilities, knowledge, and prompt creation skills through experimenting with prompts and image regeneration [Tang et al., 2024; Oppenlaender et al., 2024; Kulkarni et al., 2023]. Beyond these applications, other use cases for image regeneration include digital archiving and art restoration. For example, lost or incomplete visual assets can be recreated from a low-resolution or degraded version of the original image [Liang et al., 2023]. This technique could be utilized in preserving cultural heritage, or recreating historical imagery where traditional techniques proved difficult.

Our study builds on these motivations, emphasizing the role of the human in prompt refinement. Prior work on AI-led refinement has explored how automated systems can optimize prompts. For instance, Ma˜nas et al. presents OPT2I, a framework that leverages large language models to automatically refine prompts in txt2img models, improving alignment between prompts and generated images [Ma˜nas et al., 2024]. OPT2I iteratively revises user-provided prompts, optimizing a consistency score that evaluates how well the generated image matches the prompt, all without model fine-tuning. Similarly, a recent study by Zhan et al. on capability-aware prompt reformulation, demonstrate how refining prompt language can significantly enhance image generation quality [Zhan et al., 2024]. By tailoring prompt adjustments to match user proficiency, their system helps users create more coherent and relevant images, regardless of their familiarity with prompt engineering.

In contrast, our study emphasizes the leading role of the human in iterative refinement, investigating how individuals improve their prompts and outputs to achieve desired results. This approach aligns with prior research on human-centric systems, such as GenAssist by Huh et al. which provides blind and low-vision users with prompt-guided descriptions and visual verification features to assess generated images' content and style alignment with initial prompts [Huh et al., 2023]; but we shift the focus entirely to user-driven adjustments and improvements.

2.3 Image SimilarityMetrics

Evaluating image similarity is a fundamental task in computer vision, with applications ranging from image retrieval to quality assessment. Traditional metrics (L2 Euclidean Distance and SSIM) tend to assume pixel-wise independence, and can fall short in capturing the perceptual nuances that come from human judgment. One notable approach that addressed this limitation involves the use of deep features from convolutional neural networks (CNNs). These features are then used to assess image similarity. The Learned Perceptual Image Patch Similarity (LPIPS) metric, or simply Perceptual Similarity (PS), utilized this approach. Specifically, it computes feature embeddings from multiple convolutional layers of pre-trained networks (e.g., AlexNet, VGG, or SqueezeNet), normalizes these activations channel-wise, and then calculates the weighted L2 distance between the feature maps. The final similarity score averages these distances across spatial dimensions and layers. This approach exploits the hierarchical nature of CNNs, allowing the metric to account for higher-order image structures, context-dependent visual patterns, and other nuances that impact how humans perceive image similarity. As a result, this metric tends to align well with human judgment [Zhang et al., 2018].

Another recent approach is the Contrastive LanguageImage Pre-training (CLIP) model, which learns to associate textual and visual information. CLIP is pre-trained on largescale datasets consisting of image-text pairs, allowing it to generalize well across various domains without needing taskspecific fine-tuning. The scores produced measure the similarity between two images based on their embeddings generated by the CLIP model, without directly referencing the original text prompts [OpenAI, 2021; Wang et al., 2023].

In this study, we employ two different variants of CLIP:

B32 and L14. The B32 variant refers to a specific version with a smaller image representation size. This score is used to measure how similar two images are based on the model's understanding of their features. Similarly, CLIP L14 is another variant of the same model, but with a larger representation, capturing more detailed image features. The score indicates the degree of similarity between two images, with L14 providing a more precise detection of nuanced differences.

Finally, hashing algorithms offer an alternative approach by condensing images into compact binary representations [Buchner, 2024]. This means that the Hamming distance (the number of differing bits) between these hashes is used to indicate similarity. A smaller distance suggests greater similarity between the images, whereas a larger distance indicates more significant differences [Krawetz, 2011].

3 Research Questions &Hypotheses

This study investigates the alignment of computational image similarity metrics (ISMs) with human judgment, and subsequently how iterative prompt refinement affects image regeneration tasks. We aim to understand the relationship between iterative prompt refinement and the tools used to measure its effectiveness. The following research questions guide our investigation:

RQ1 - Human Perception of ISMs: Do humans generally agree that the selected ISMs are reliable numerical heuristics for evaluating whether two images are perceived as similar or different?

While Zhang [Zhang et al., 2018] shows that their proposed metric, Perceptual Similarity, outperforms previous traditional metrics (SSIM, FSIM, L1 or L2 norm) in aligning with human judgments, Sinha and Russell [Sinha and Russell, 2011] demonstrate the limitations of ISMs, cautioning users when interpreting their reliability. This raises the question of whether humans consistently perceive these metrics as reliable indicators of similarity. Addressing RQ1 supports our study by assessing how humans perceive ISMs as tools to help facilitate iterative refinement in image regeneration tasks. The hypotheses for RQ1 are then as follows:

• Hypothesis 1.1: Human raters exhibit moderate to good agreement with the ISMs as evaluative tools for image similarity.
• Hypothesis 1.2: There are no significant differences across the ISMs in terms of agreement with subjective human assessment.

RQ2 - Impact of Iterative Prompt Inference on Image Regeneration: Given the task of prompt inference to regenerate a target image, does iterative prompt inference improve the ISM score of a user-generated image, meaning it is more similar to the target image?

Building on prior work demonstrating the potential of iterative prompt learning to enhance image alignment with target outputs [Liang et al., 2023; Ma˜nas et al., 2024; Zhan et al., 2024], we investigate RQ2 in order to determine whether iterative refinement of user-generated prompts results in generated images that more closely match target images as measured by ISMs. The following are hypotheses tied to this research question:

• Hypothesis 2.1: Iterative prompt refinement improves the similarity of user-generated images with target images.
• Hypothesis 2.2: Improvement in ISM scores diminishes over successive iterations.
• Hypothesis 2.3: Users perceive images generated from later iterations as more similar with target images compared to earlier iterations.

The tests conducted to evaluate these hypotheses are further explained in Section 6.

4 Survey Design

4.1 Task: Image Regeneration through Iterative Prompt Refinement

Each participant was assigned a set of 10 target images, chosen from the image dataset, and tasked with refining prompts over 10 iterations per image, in order to reproduce the target image. For half of the target images, ISM feedback was displayed to participants, while the remaining half omitted this feedback to assess its influence on performance. To ensure the chosen ISMs are tested equally, our participant pool of 20 was divided into 4 subsets of 5, each subset testing a different image similarity metric.

4.2 Task: Subjective Similarity Ranking

Following the iterative refinement task, participants ranked the 10 images generated across iterations for each target image. Rankings were performed using a drag-and-drop interface, with the most similar images placed on the left and the least similar on the right.

5 Mixed-Effects Model

To model both repeated measurements and hierarchical structure, we fit a mixed-effects model with fixed effects (e.g., iteration, demographics, txt2img familiarity) and random intercepts for each participant (session id) and each target prompt [Moerbeek, 2004]. Each participant, identified by a unique session id, completed multiple target prompts, with 10 iterations per prompt. This design yields nested repeated measures: iterations within prompts, and prompts within participants.

We modeled random intercepts for both session id and target prompt to account for between-subject and between-prompt variability. To capture temporal autocorrelation within each prompt iteration sequence, we applied an AR(1) covariance structure on residuals, assuming stronger correlation between temporally adjacent iterations.

The adjusted ISM score was modeled as:

Random Effects:

• bj ∼ N(0, σ2 b,j ): session-level intercept
• bk ∼ N(0, σ2 b,k ): prompt-level intercept

Residuals:

where ρ captures autocorrelation between iterations.

Estimation: Restricted maximum likelihood (REML).

6 Experimental Setup &Evaluation

6.1 RQ1: Evaluating Alignment

Although our primary focus is ultimately on the human assessment of image quality, we begin by examining whether the ISMs can approximate human perceptions of similarity to aid in quantitative analysis. In our dataset, each target prompt was rated by a human and by a given ISM. By comparing these sets of ranks, we investigate whether an ISM's ordering of images correlates meaningfully with how humans order them. If an ISM score aligns with human judgments, it may serve as a useful heuristic in identifying which images are more (or less) similar to the target. However, it is important to approach these results with caution, as no automated metric can fully capture the nuanced ways in which humans evaluate images.

Intraclass Correlation Coefficient: To evaluate alignment of the chosen ISMs with subjective human ratings, we employed the Intraclass Correlation Coefficient (ICC). This metric is specifically designed to assess the degree of agreement or consistency among raters who evaluate the same set of items. Unlike the Pearson correlation, the ICC takes into account not only the linear relationship but also the consistency in how items are scored. In our study, we used a two-way mixed-effects model with a consistency definition, treating the ISM ranking as a fixed effect (since it is a specific algorithm whose performance we want to evaluate) and the images as random effects. This model is suitable because we are interested in whether an ISM's relative ordering of items parallels that of human raters, rather than whether the machine matches human scores exactly.

We interpret the ICC using conventional guidelines for reliability [Koo and Li, 2016]: values below 0.5 are poor, values between 0.5 and 0.75 are moderate, values between 0.75 and 0.90 are good, and values greater than 0.90 are excellent. An unacceptably low ICC value would suggest that an ISM has little resemblance to the human rankings (i.e., the metric is not a good heuristic), whereas a moderately high or better ICC suggests that the ISM reflects a useful, though still imperfect, approximation of human perceptions.

For our analysis, we computed ICC values for different metric types to determine which ones align most closely with human judgments. Table 1 summarizes the ICC results for the four ISMs we evaluated. All metrics exhibit statistically significant ICC values at p < .001, indicating that each metric aligns with human ratings at levels significantly above zero agreement; however, their degree of alignment significantly varies.

We found that B32, L14, and Perceptual Similarity (PS) attained moderate agreement with human raters. Their respective ICC values exceed 0.50, suggesting that while they may not perfectly replicate human judgments, they sufficiently capture a meaningful portion of how humans perceive image similarity for the purpose of this study. Consequently, these metrics can reasonably serve as proxies for human preferences in subsequent analyses.

In contrast, ImageHash yielded a notably lower ICC of 0.250, signifying poor alignment with human evaluations. Because our aim is to ensure that each ISM used in the study reflects human judgments to an acceptable degree, we have decided to exclude ImageHash from the next stage of analysis. By removing ImageHash, we focus our analyses on those metrics that offer more credible approximations of humandriven perceptions of image similarity.

The ICC values in Table 1 demonstrate that B32, L14, and PS achieve moderate alignment with human judgments, which supports Hypothesis 1.1. However, the notably lower ICC for ImageHash (0.250) suggests that not all metrics yield equivalent alignment. Thus, Hypothesis 1.2, predicting no significant differences among the ISMs, holds only for B32, L14, and PS, but not for ImageHash. This is why ImageHash was excluded from further modeling, ensuring we focus subsequent analyses on the ISMs that fulfill Hypothesis 1.1's criterion of moderately capturing human judgments.

6.2 RQ2: Evaluating Refinement

Mixed-EffectsModel Results: After excluding the ImageHash metric, we fit a linear mixed-effects model, defined in Section 5 to examine how our fixed factors and random intercepts contribute to the adjusted ISM. Table 2 presents the Type III Tests of Fixed Effects. We observe that:

• Iteration is significant. We observe (F(9, 1451) = 11.486, p < .001), suggesting that the adjusted score

changes systematically across successive iterations. Posthoc comparisons of the iteration estimates (Table 2) indicate an overall trend toward improved adjusted scores over the first several iterations, with diminishing effects after iteration 6. From Table 3, we see that iterations 1 through 6 each exhibit significantly lower (improved) adjusted scores compared to the reference level (iteration 10). Specifically, iteration 1 has the largest negative estimate (-0.053), and although the magnitude of improvement tapers with increasing iteration number, all estimates remain significantly different from zero through iteration 6. By iteration 7 and beyond, the effect is no longer statistically significant, implying that user performance begins to plateau around the seventh iteration.

• Subject, the content of the prompt, (e.g., "cat," "astronaut," etc.) also shows a significant main effect (F(4, 1451) = 5.758, p < .001). This indicates that some subjects inherently tend to yield higher or lower adjusted scores, irrespective of other predictors.
• Visibility of ISM is not significant (F(1, 1451) = 2.061, p = .151), nor is the interaction between iteration and visibility of metric (F(9, 1451) = 0.515, p = .865). Thus, showing the metric during the task does not reliably alter the rate of improvement across iterations in this dataset.
• The specific type of ISM (PS, CLIP L14, or CLIP B32) shown to a user does not exhibit a statistically significant main effect under this model (F(2, 1451) = 0.446, p = .504). Given that we already established all three to have acceptable alignment with human rankings via the ICC, this result suggests that once the model controls for other factors, the three ISMs produce broadly similar adjusted scores on average.

Table 4 shows the estimates of the random effects and the AR(1) correlation in the residuals. We included random intercepts for session id and target prompt to account for unexplained variability at both the participant and prompt level. Both variance components are small but highly significant (p < .001), indicating that individual differences between sessions and systematic differences between prompts do exist. Additionally, the AR(1) correlation ρ = -0.230 is statistically significant (p < .001). Although negative serial correlation may seem counterintuitive, it indicates that if a participant's adjusted score is above the model's prediction at one iteration, it tends to be slightly below the model's prediction at the next iteration (and vice versa).

Overall, these results suggest that iteration and subject have robust influences on adjusted scores, whereas the visibility of metric and the type of metric do not produce strong differential effects. The random-effects estimates affirm that allowing each participant and each target prompt to vary with its own intercept meaningfully improves model fit. These findings inform our subsequent interpretations of user performance. Because iteration consistently emerges as a key predictor, our data suggest that participants' scores improve over time. Meanwhile, the negative AR(1) coefficient indicates small but significant oscillations from iteration to iteration in how users respond.

The mixed-effects model results highlight that Iteration emerged as a significant predictor of the adjusted ISM score, with iterations 1 through 6 each showing improved scores relative to iteration 10. This directly supports Hypothesis 2.1: successive iterations lead to meaningful gains in similarity with the target image. Moreover, these improvements diminish beyond iteration 6, suggesting a plateau effect consistent with Hypothesis 2.2. Taken together, these patterns indicate that most of the iterative benefit is realized in the earlier cycles of prompt refinement, after which further iterations yield less significant enhancements.

Top User-Ranked Images: While the mixed-effects model in the Section 5 relies on ISM-based scores, we also examined a purely human-centric measure of iterative improvement. Specifically, from the ranking procedure explained in the Section 4.2, we evaluated from which iteration came a user's top-ranked image. If iterative prompting has no effect on the user's perception of which generated image is closest to the target image, we would expect that a user's top-ranked image is equally likely to come from any of the 10 iterations.

To test this, we aggregated the iteration at which each user selected their top-ranked image into one distribution and performed a chi-square goodness-of-fit test against the null hypothesis of a uniform distribution. Table 5 shows the observed and expected frequencies per iteration. The result was highly significant, χ2 (9) = 71.200, p < .001, indicating that users most frequently selected images generated in later iterations (particularly 9 and 10), rather than early ones. In other words, the best image rarely appeared in the initial iterations, providing additional evidence that users iteratively refine their prompts over time to achieve better results.

This human-only perspective, independent of ISM scores, provides evidence and support for Hypothesis 2.3, that iterative prompting contributes to users' subjective sense of improvement in the image regeneration task. Users predominantly favored images from their final two iterations.

7 Discussion

Implications from Results: Generative AI has rapidly expanded, making prompt engineering an increasingly critical skill [Oppenlaender et al., 2024]. Our findings confirm that single-shot prompting often falls short when aiming for a precise target visual; in contrast, iterative prompt refinement offers a more positive path to alignment. From the results of our analysis in Section 6; the following insights stand out:

• Alignment between subjective human assessment and objective ISM scores: Moderate agreement between human assessment and ISMs, as demonstrated by the ICC values in Section 6.1, demonstrates that ISMs like Perceptual Similarity (PS) and CLIP variants can reasonably approximate human perception. This is further supported by research such as Ghildyal and Liu, which demonstrated a new metric based on PS that is robust to small misalignments in aligning with human perception [Ghildyal and Liu, 2022]. Zhang and Krawetz have shown that that pixel-wise image comparison tend to not align well with human perception, which correlates to ImageHash's ICC being the lowest in our results [Zhang et al., 2018; Krawetz, 2011]. Overall, this has potentials for the development of educational tools, where these metrics could guide novice users in refining prompts to achieve a desired AI-generated images.
• Effectiveness of iterative prompt refinement: The results from the mixed-effects model in Section 6.2 serves as an objective demonstration that iterative refinement meaningfully improves the alignment of user-generated images with target images, as evidenced by the improvement in

ISM over the first six iterations, followed by a plateau in score until the last iteration. This plateau could suggest that participants show the most improvement in the earlier iterations, with a potential for a diminishing return in the later iterations. Moreover, users predominantly favor images from their final two iterations, meaning there is a subjective sense of improvement in their image regeneration task. Overall, the results confirms the utility of iterative approaches of human when performing generative AI-related tasks; but also highlight a potential for further research on enhanced guidance or feedback support for users to overcome the improvement plateau. Since previous research has shown effectiveness in AI-led prompt refinement [Ma˜nas et al., 2024; Zhan et al., 2024; Liang et al., 2023]; our results can potentially open up avenues in research for human-AI collaboration in iterative prompt refinement.

Limitations and Future Work: While our results are promising, several limitations warrant discussion. First, our sample of 20 university students was small and homogeneous; future work should recruit more diverse populations to reveal broader patterns. Second, the fixed number of iterations per target image may not fully capture the potential of iterative refinement, and may have introduced biases in later iterations, as seen in top user-ranked results (Section 6.2). Future studies could explore flexible iteration counts or determine where performance gains plateau.

Third, although we controlled for variation in target prompt conditions, assigning identical prompts to all participants could better reveal how prompt content affects ISM scores (Section 6.2). We also found ISM feedback alone had minimal impact, highlighting the need for more intuitive feedback mechanisms.

Our focus on txt2img models leaves open whether similar effects hold in other domains like language, code, or audio generation. Research into these areas could validate the generalizability of our findings. Finally, as iterative refinement enhances creative capacity, it also raises ethical issues: misinformation, deepfakes, and IP rights [OECD.AI, 2025; Marcus and Southen, 2024]. Future work should address these concerns via training, policy, or collaborative safeguards.

Conclusion

Our study demonstrated that iterative prompt refinement substantially enhances alignment between AI-generated images and target visuals, particularly in the early stages of refinement. Moderate agreement between select ISMs and human evaluations indicated that such metrics hold promise as objective feedback tools for user workflows. Nevertheless, limitations related to participant diversity, iteration count, and feedback mechanisms point to the need for further research. Given the increasing prevalence of AI-generated content on social media platforms as well as the web as a whole, this work provides a solid understanding for optimizing humancentric workflows in generative AI tasks, noting the importance of iterative refinement for achieving desired results.