acceptodds
Under review as a conference paper at ICLR 2027

ModEval: Benchmarking Steering Modifiers for Social-Risk Evaluation in Text-to-Image Generation

Abstract

Text-to-image (T2I) models have been widely adopted for their impressive ability to generate high-quality images through textual prompts. During prompt composing, users often rely on prompt modifiers, i.e., auxiliary descriptions attached to the subject, to enhance visual richness and control artistic form. While modifiers provide a lightweight interface for steering generation, they may also alter model behavior in ways that lead to problematic outputs, raising broader social concerns. In this work, we propose ModEval, a unified evaluation framework for benchmarking the social impact of steering modifiers in T2I generation. We first harmonize modifier taxonomies into eight categories and construct the Prompt Modifier Dataset (PMD), which contains approximately 74,000 real-world prompts annotated with span-level modifier categories. Using PMD, we identify five steering modifier categories through dual-metric validation, combining image-level and semantic-level influence measurements. We then evaluate these steering categories across three social dimensions: demographic bias, deepfake detectability, and safety, on both open-source and black-box commercial T2I models. Our results show that all five steering categories can induce downstream social risks with different emphasis. Artist modifiers are primarily associated with demographic bias shifts; trending and medium modifiers substantially affect deepfake detectability; and atmosphere and movement exhibit broader multi-risk effects, with atmosphere most strongly linked to safety concerns.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
text-to-image model, prompt modifier, social risk evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4jMnOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how prompt modifiers, such as artist names, styles or mood words, affect social risks in t2i models. The authors build PMD dataset of real prompts with labeled modifiers and try to identify which modifier types change the output. They then test those modifier types for effects on gender and race representation, how easily fake-image detectors flag the images, and on unsafe content.

Critical strengths
  1. An interesting perspective on how the prompt modifiers, such as artist names, styles or mood words, affect social risks in t2i models.
  2. The released dataset of annotated real-world prompts should be useful to the community.
  3. The evaluation covers diverse dimensions, such as demographic bias, fake-image detectability, and content safety, providing a broader view of how modifier effects.
Critical weaknesses
  1. Novelty relative to prior work. [1] already show that modifiers can unlock unsafe generation, and [2] study modifiers and bias. The added contribution seems to be the PMD dataset and the category-level organization.

  2. FairFace classifiers and the HuggingFace deepfake detector are trained on photographic faces or images. The independent variable here is style. Classifier errors are therefore likely to correlate with the modifiers in the study.

  3. Safety spans come from MMD, e.g., "sinister", "horror art", "Beksiński", "medical illustration" (Fig. 12). It is not surprising that "sinister" raises Q16 gore flags on "a person." Thus the safety experiment may be confounded by how spans were chosen.

[1] Modifier Unlocked: Jailbreaking Text-to-Image Models Through Prompts (2025)

[2] Can Prompt Modifiers Control Bias? A Comparative Analysis of Text-to-Image Generative Models (2024)

Questions for authors
  1. What is the main contribution beyond [1, 2] ? (weakness1)
  2. How modifiers affect the classifier? (Weakness2)
3 comments
AuthorsOpenReview ↗

Response to Reviewer 4jMn (Part 1)

We thank the reviewer for the weak acceptance rating, and for the insightful and constructive comments. We are confident to address all the concerns raised in weaknesses and questions.

W1, Q1. Contribution beyond modifier-based jailbreaking and modifier-controlled bias.

We clarify that the open question ModEval answers is which categories of prompt modifiers steer T2I generation, and which of these categories drive which social risks. Prior work, including [1] and [2], treats modifiers as a single, undifferentiated prompting tool. ModEval instead analyzes them at the category level. PMD is also a new modifier dataset we constructed for future modifier and T2I model research.

  • The steering categories. ModEval is the first to test which modifier categories steer generation. We harmonize existing taxonomies into a unified set of categories and build PMD, the first span-level dataset of real-world prompt modifiers with category labels. We then propose a dual-metric validation that separates categories which genuinely steer generation from those which are frequently used but have little influence. This category-level distinction is a blind spot of [1] and [2], and it determines which modifiers are worth evaluating for social risk in the first place.
  • Steering effect within and across risks. [1] studies safety only, and [2] studies bias only. ModEval evaluates the same steering categories under one protocol across three risk dimensions, which enables us to compare categories within and across risks, and shows that each category has its own risk profile. Such profiles can motivate category-aware safeguards instead of general filtering.

Novelty over modifier-based jailbreak attack. [1] shows that modifiers can unlock unsafe generation, but it assumes an attacker who deliberately crafts or searches for modifiers that bypass safety filters. ModEval DOES NOT require an attacker. Even a benign user can generate unsafe images by accident, because many unsafe-associated modifiers are ordinary style words such as "eerie" or "ominous". Our question is therefore not whether a filter can be bypassed. Instead, we ask which steering categories make unsafe outcomes more likely. Beyond safety, we also analyze other social-risk dimensions, which enables a cross-risk analysis.

Novelty over modifier-controlled bias. [2] asks whether prompt modifiers can control bias. It treats modifiers as an intentional tool for adjusting demographic outcomes. ModEval asks the opposite question. We study ordinary style modifiers that users add for aesthetic reasons. These modifiers contain NO demographic words, and we ask whether they still change demographic representation as a side effect. Our bias modifiers are the most frequent spans of each steering category in real-world prompts, rather than spans selected to control bias. Beyond bias, we also analyze other social-risk dimensions, which enables a cross-risk analysis.

[1] Modifier Unlocked: Jailbreaking Text-to-Image Models Through Prompts (Liu et al., IEEE S&P 2025)

[2] Can Prompt Modifiers Control Bias? A Comparative Analysis of Text-to-Image Generative Models (Shin et al., IJCAI Workshop 2024)

AuthorsOpenReview ↗

Response to Reviewer 4jMn (Part 2)

W2, Q2. Classifier errors and the effect of the modifiers on classifiers.

We agree that classifier errors may correlate with style. We therefore measure how modifiers affect each classifier directly.

FairFace classifiers. We validate them against human labels on a stratified sample of 460 images: the 100 images in Section 5 plus 360 new images (20 per model × group), covering the baseline and all five steering categories on the three models. Two authors independently label perceived gender and race. Agreement is 96.7% for gender (κ = 0.94) and 92.0% for race (κ = 0.82), and disagreements are resolved by discussion. The table gives the accuracy of each classifier against the consensus labels.

ClassifierBaselineArtistAtmosphereMediumMovementTrending
Gender (FairFace Gender Image Detection)0.950.920.910.970.880.93
Race (fairface-race-vit)0.920.820.910.840.860.92
  • Gender. Accuracy stays between 0.88 and 0.97 in every category, close to the baseline (0.95).
  • Race. Accuracy is 0.92 at baseline and 0.82–0.92 with modifiers. Modifiers lower it by at most 0.10, and every category stays above 0.80.
  • Classifier errors cannot explain the shifts. The shifts in Section 4.2 are much larger than these error rates. For example, "a CEO, alphonse mucha" on FLUX.1-dev turns 100% men into 100% women, while at most 8% of artist images are misclassified for gender. Modifiers therefore affect the FairFace classifiers only moderately. They are reliable for gender and race.

Deepfake detector. We also agree with the reviewer that the original detector is trained on face images. In our revised version, we apply it to person images only, which lie within its training domain. We further add a second detector, Organika SDXL-detector, for both person and street images. It is fine-tuned from an AI-art detector trained on images in many artistic styles, and its real images cover general scenes rather than faces only. To see how modifiers affect detection, we label each image as photorealistic or stylized with CLIP zero-shot classification (88% agreement with human labels on 156 images). We then fit a logistic regression of detection on category and model, with and without a photorealism indicator.

  • Modifiers mainly change style. Except for atmosphere, they lower the photorealistic share of portraits from 79% to 5–9%.
  • The original detector responds to this style change. All five categories significantly lower its detection of portraits, but none remains significant once photorealism is controlled.
  • Category effects remain under the second detector. After controlling for photorealism, artist and medium still lower the detection of portraits (coefficient −1.19, p = 0.004), and atmosphere, movement and trending still raise the detection of street images (+0.58 to +0.63, all p < 0.05).

Modifiers therefore affect the original detector mainly through style. The second detector covers stylized images, and its category effects remain after controlling for style. The category-level effects on detectability are therefore not explained by detector failure on stylized images.

We will report the 460-image validation with the per-category table, and add the second detector and the photorealism analysis to our revised paper.

AuthorsOpenReview ↗

Response to Reviewer 4jMn (Part 3)

W3. Choice of safety spans.

Our goal in the safety experiment is to compare how modifier categories affect unsafe generation. To measure this effect, we deliberately choose modifiers that can trigger unsafe outputs. This is because common modifiers rarely trigger them. With the five most frequent spans of each category in PMD, the gore rate is 10% (13% at baseline) and the explicit rate is 1.5% (0% at baseline). These spans can cause a small amount of unsafe generation, but they cannot give a convincing comparison between categories.

We therefore use unsafe-associated spans from MMD. This is not jailbreak prompt design. We do not craft or optimize prompts. We only group the released spans by category, and the spans of every category are chosen in the same way. We agree that a span such as "sinister" can raise gore flags by itself. However, this does not explain the differences between categories, which are the focus of our analysis.

Gore rate pooled over the three models:

CategoryFrequent spans (PMD)Unsafe-associated spans (MMD)
Baseline12.7%–
Artist8.7%42.0%
Atmosphere10.7%95.3%
Medium10.0%30.5%
Movement6.7%74.7%
Trending14.0%14.7%
  • The same selection gives very different outcomes. Under unsafe-associated spans, the gore rate ranges from 15% (trending) to 95% (atmosphere). "Beksinski" (artist) and "medical illustration" (medium) are also unsafe-associated, yet these two categories produce much less gore (42% and 31%). If the span selection drove the result, all categories would rise in a similar way.
  • Category effects remain after controlling for span wording. We score each span by its CLIP text similarity to "gore, violence, blood" and add this score to a logistic regression of gore on category and model (artist as reference). Atmosphere remains significantly higher (coefficient +1.09, p < 0.001), and medium and trending remain significantly lower (p ≤ 0.008). The span score itself has no significant effect (p = 0.64). For explicit content, the category differences also remain significant after this control.

The category ranking therefore reflects the modifier categories rather than how the spans were chosen. In the revised paper, we will add this analysis.

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

The paper studies how prompt modifiers affect demographic representation, synthetic-image detectability and safety in text-to-image generation. Existing modifier taxonomies are harmonized, and an LLM annotates modifier spans in 73,718 real-world prompts to build the Prompt Modifier Dataset. Image-similarity and semantic-alignment tests are then used to select five steering modifier categories. These are evaluated on FLUX.1-dev, SD3.5-Large and Midjourney v7.

Critical strengths
  1. PMD looks reusable: it adds span-level categories and character offsets to a large prompt corpus, and the LLM annotator is checked against a human-labeled reference set.

  2. Three models are compared on demographic representation, detectability and safety under one taxonomy. Some observations are counterintuitive, such as detection rates moving in opposite directions for the open-source models and Midjourney.

Critical weaknesses
  1. The framework largely combines existing taxonomies, LLM annotation and off-the-shelf classifiers, and it is not clear which open question the new annotations and evaluation dimensions answer.

  2. The analysis stays descriptive. The opposite detection trends, for instance, are reported but not separated into changes in realism, style or detector sensitivity, so the category-level conclusions are hard to generalize.

  3. Some measured quantities sit far from the claimed social risk: a shifted demographic distribution is not automatically more bias and a lower detection rate is not automatically more deceptive imagery. Q16 is a general inappropriate-content detector used here for gore without task-specific validation, and the cross-dimension synthesis takes absolute changes, mixing risk increases and decreases into one profile.

Questions for authors
  1. What is the most important new finding relative to prior work on modifier-induced bias and modifier-based jailbreaking? Can the existing images distinguish a style change from a realism change or a detector failure in the opposing detection trends?

  2. Could you report human validation broken down by model, modifier category and explicit-content versus gore, in particular a confusion matrix for Q16 against human gore labels?

  3. Do the category rankings and cross-dimension conclusions survive if increases and decreases are treated separately? For demographics, can you separate a departure from the baseline distribution from a deterioration under a stated fairness criterion?

5 comments
AuthorsOpenReview ↗

Response to Reviewer HRkF (Part 1)

We thank the reviewer for the constructive and insightful comments. We are confident that we can address the concerns point to point and revise the paper accordingly.

W1, Q1 (first question). Research Question and Most Important New Findings.

The open question ModEval answers is which categories of prompt modifiers steer T2I generation, and which of these categories drive which social risks. Prior work treats modifiers as a single, undifferentiated prompting tool. ModEval instead analyzes them at the category level.

  • The steering categories. ModEval is the first to test this assumption at the category level. We harmonize existing taxonomies into a unified set of categories and build PMD, the first span-level dataset of real-world prompt modifiers with category labels. We then propose a dual-metric validation that separates categories which genuinely steer generation from those which are frequently used but have little influence. This category-level distinction is a blind spot of prior work, and it determines which modifiers are worth evaluating for social risk in the first place.
  • Steering effect within and across risks. ModEval evaluates the same steering categories under one protocol across three risk dimensions, which enable us to compare categories within and across risks, and shows that each category has its own risk profile. Such profiles can motivate category-aware safeguards instead of general filtering.

Novelty over modifier-controlled bias. Existing work on modifiers and bias asks whether prompt modifiers can control bias. It treats modifiers as an intentional tool for adjusting demographic outcomes. ModEval asks the opposite question. We study ordinary style modifiers that users add for aesthetic reasons. These modifiers contain no demographic words, and we ask whether they still change demographic representation as a side effect. Our bias modifiers are the most frequent spans of each steering category in real-world prompts, rather than spans selected to induce bias. Beyond bias, we also analyze other social-risk dimensions, which enables a cross-risk analysis.

Novelty over modifier-based jailbreaking. Modifier-based Jailbreaking studies assume an attacker who deliberately crafts or searches for modifiers that bypass safety filters. ModEval does not require an attacker. Even a benign user can generate unsafe images by accident, because many unsafe-associated modifiers are ordinary style words such as “eerie” or “ominous”. Our question is therefore not whether a filter can be bypassed. Instead, we ask which steering categories make unsafe outcomes more likely. Beyond safety, we also analyze other social-risk dimensions, which enables a cross-risk analysis.

W2, Q1 (second question). Generalization of category-level conclusions. We agree that the original analysis is a bit descriptive. The existing images can separate these factors once each image carries a style label. We therefore add three analyses.

  1. We label every image as photorealistic or stylized with CLIP zero-shot classification, which agrees with our human labels on 88% of 156 annotated images.
  2. We add a second detector, Organika SDXL-detector, whose training data cover general scenes and diffusion-generated images rather than faces only. This can help mitigate the single detector error.
  3. We fit a logistic regression of detection on category and model, with and without a photorealism indicator.

We obtain the following results.

TermOriginal, person: withoutOriginal, person: withOrganika, person: withoutOrganika, person: withOrganika, street: withoutOrganika, street: with
artist-0.55 (.020)+0.13 (.629)-1.54 (<.001)-1.19 (.004)+0.28 (.264)+0.04 (.877)
medium-0.83 (<.001)-0.11 (.683)-1.54 (<.001)-1.19 (.004)+0.49 (.048)+0.25 (.346)
movement-0.75 (.002)-0.06 (.827)-0.54 (.175)-0.19 (.675)+0.87 (<.001)+0.58 (.034)
trending-1.01 (<.001)-0.29 (.298)-0.84 (.030)-0.49 (.253)+0.81 (.001)+0.63 (.016)
atmosphere-0.61 (.010)-0.37 (.135)-0.33 (.417)-0.20 (.625)+0.68 (.007)+0.62 (.013)
photorealistic–+0.97 (<.001)–+0.53 (.086)–-0.48 (.010)

Please see detailed explanation regarding Realism, style, Detector sensitivity, and Category effects beyond style below in Response to Reviewer HRkF (Part 2).

AuthorsOpenReview ↗

Response to Reviewer HRkF (Part 2)

  • Realism. Modifiers make images less realistic. For person images, the photorealistic share at baseline is 94% (FLUX.1-dev), 86% (SD-3.5-Large) and 58% (Midjourney v7), and it falls to less than 18% with artist, medium, movement and trending. Atmosphere keeps about half of its outputs photorealistic.
  • Style. With the original detector, all five categories significantly lower the detection of portraits, but none remains significant once photorealism is controlled. For this detector, the decrease reflects the shift from photorealistic to stylized outputs.
  • Detector sensitivity. Detectors differ in how sensitive they are to each generator, and this explains the opposite trend on Midjourney v7. The original detector is much less sensitive to photorealistic Midjourney v7 portraits (21%, vs. 81% for FLUX.1-dev and 74% for SD-3.5-Large), while it responds to stylized outputs similarly for all models. Stylization therefore moves the three models in opposite directions under this detector. This is why we add a second detector: under it, the increase on Midjourney v7 disappears (ΔDDR from −0.28 to +0.02), while the category effects in the table remain.
  • Category effects beyond style. Under the second detector, artist and medium still lower the detection of portraits, and atmosphere, movement and trending still raise the detection of street images after controlling for photorealism (Table B). These effects are not explained by a change in realism, and we report them as the category-level findings that generalize.

W3, Q2. Human validation of Q16.

We agree with the reviewer that Q16 needs task-specific validation. We therefore conduct human validation. Two authors independently label a stratified sample of 240 safety images as safe, explicit or gore. The sample covers the three models, both scenarios (explicit and gore prompts), the baseline and all five steering categories.

The iInter-annotator agreement is κ = 0.80 (89.6%), and we use the 215 images on which both annotators agree. We report the results by model, by modifier category and by content type, followed by the confusion matrix of Q16 against the human gore labels.

  • By model. Q16 recall on gore is 0.92, 1.00 and 1.00 on FLUX.1-dev, SD-3.5-Large and Midjourney v7. On every model, atmosphere and movement have the two highest human gore rates. For explicit prompts, humans label 32% of FLUX.1-dev and 38% of SD-3.5-Large images as explicit, but only 7% of Midjourney v7 images. This matches Section 4.4: explicit content appears mainly on the open-source models.
  • By modifier category. Human labels give the same category ranking as Section 4.4. Under gore prompts, humans label 84% of atmosphere and 78% of movement images as gore, against 0–10% for the other categories and the baseline. Under explicit prompts, medium has the highest human explicit rate (53%), followed by trending (42%) and atmosphere (21%). The baseline is 0%. Q16 recall is 0.95–1.00 in every category with human gore images, and its precision is highest on atmosphere and movement (0.87 and 0.72).
  • Explicit content vs. gore. For gore, Q16 finds 41 of the 42 human gore images (recall 0.98). For explicit content, every image NudeNet flags is explicit to humans (precision 1.00).

Confusion matrix of Q16 against human gore labels (215 images):

Human: goreHuman: not goreTotal
Q16: gore412263
Q16: not gore1151152
Total42173215

We will add this validation in our revised version.

AuthorsOpenReview ↗

Response to Reviewer HRkF (Part 3)

W3, Q3 (first question). Separating increases from decreases in risk profile.

Our original design does not consider the direction of change. This is because we treat both an increase and a decrease of risk as an effect of the modifier on that risk dimension. We thank the reviewer for this point. We agree that treating the two directions separately makes the risk assessment more meaningful. We therefore recompute Figure 4 for increases and decreases separately. A risk increase means a larger gender or race imbalance for bias, a lower detection rate for detectability, and a higher NSFW rate for safety. A risk decrease is the opposite. For detectability, we use the original detector on person images and Organika on street images. All other steps follow Appendix G.1. We obtain the following results.

Increase-only profile:

CategoryBiasDetectabilitySafety
Artist45.2%35.5%19.2%
Medium41.0%39.6%19.4%
Movement34.5%33.7%31.9%
Trending41.9%48.1%10.0%
Atmosphere27.4%24.8%47.9%
Decrease-only profile:
CategoryBiasDetectabilitySafety
------------
Artist56.4%43.6%0.0%
Medium35.6%38.3%26.1%
Movement52.9%47.1%0.0%
Trending23.9%38.1%38.1%
Atmosphere61.0%39.0%0.0%
  • Risk increases. The strongest category in each dimension stays the same: artist for bias, trending for detectability and atmosphere for safety. The safety ranking is identical to the paper. Artist keeps the largest bias share (45.2%). Trending and medium keep the two largest detectability shares (48.1% and 39.6%). Atmosphere keeps the largest safety share (47.9%). Movement remains the most balanced profile (31.9–34.5% on all three dimensions).
  • Risk decreases. Safety decreases are almost zero (at most 0.03, against up to 0.46 for increases). Modifiers rarely make outputs safer, so the safety shares in the decrease table come from very small values. Bias decreases are largest for movement and atmosphere, where some occupations move closer to parity. Trending has the largest detectability change in both directions.

Therefore, the category rankings and cross-dimension conclusions survive when increases and decreases are treated separately. We will add both profiles to our revised version.

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

The paper introduces ModEval, a framework for evaluating how prompt modifiers influence t2i generation and associated social risks. The authors first consolidate existing modifier taxonomies into eight categories and construct the Prompt Modifier Dataset (PMD), containing approximately 74,000 real-world prompts annotated with modifier spans using an LLM. They then identify five steering categories (artist, medium, movement, trending, and atmosphere). These categories are subsequently evaluated across three dimensions: demographic bias, deepfake detectability, and safety. The results suggest that artist modifiers have the strongest effects on gender representation, medium and trending modifiers substantially affect synthetic-image detectability, and atmosphere and movement modifiers are associated with higher unsafe-content detection rates.

Critical strengths
  1. I find the question of the paper relevant. Studying the prompt effectrs at the modifier level, rather than treating prompts as a whole, is an interesting direction.
  2. The dataset the authors construct could be a useful resource for future research on prompting and model behavior.
  3. The paper is generally well organized around two research questions, and the authors also include human validation of the automated annotations and classifiers.
Critical weaknesses
  1. The selection of steering categories (rq1) is based on a very small experiment (25 prompts with 4 images each), and also the effects are quite small. The pass/fail decisions are also borderline in several cases. Since the rest of the paper builds on this step, a comparison with a non-steering category (e.g., quality) in rq2 would test whether the rq1 selection matters.

  2. For the gender and racial bias experiment, I am not sure whether the fairface classifiers are reliable, given that they have been trained on real-world face photographs. It is unclear whether they perform equally well on images generated with artistic modifiers (e.g. paintings). The authors have performed a small validation of 100 images, could they perhaps also report accuracy per modifier category?

  3. The paper measures how much modifiers change demographic distributions, but its conclusions do not distinguish changes that increase imbalance from changes that reduce it, even though Table 2 shows both. Labelling all such shifts as “bias” therefore overstates the risk. Could the authors distinguish between modifiers that increase and decrease demographic imbalance, and clarify how the observed distributional shifts should be interpreted as social risks?

  4. The deepfake detectability evaluation relies on a single detector. According to its model card, it was fine-tuned on real face photographs and deepfake face images, and the card itself notes that it may not generalize to other domains. This is particularly concerning because the evaluation includes city streets (entirely outside the training domain) and images in non-photographic artistic styles. Could the authors evaluate at least one additional detector?

  5. Apart from concerns about the reliability of the deepfake detector, I am not convinced that stylized images necessarily pose a deepfake risk. Modifiers such as “pixiv”, “digital art”, or “sci-fi” make images look like illustrations (see Figure 11), and a detector trained on photographs may simply struggle to classify these images correctly. However, I don't think someone would mistake an anime drawing for a real photograph. A real deepfake risk would involve realistic images that can actually fool the detector. Could the authors report results separately for photorealistic and stylized outputs?

Questions for authors

Please see questions posed in the "Critical Weaknesses" section.

8 comments
AuthorsOpenReview ↗

Response to Reviewer QGVG (Part 1)

We thank the reviewer for the constructive and insightful comments. We are confident that we can address the concerns raised in the critical weaknesses point to point and are willing to revise the paper accordingly.

W1. Robustness of the RQ1 steering selection: sample size, borderline effects, and a non-steering control.

We agree that 25 prompts with 4 images each is a relatively small basis for a step the rest of the paper builds on. We therefore enlarge RQ1 to 100 prompts per category. To address effect size and borderline decisions directly, we now report three quantities next to each p-value. We also run Quality through all three RQ2 dimensions as a negative control to make our results more convincing. Our new evaluation supports the original selection of five steering categories, and we report our new analysis in detail.

RQ1 enlarged to 100 prompts. We draw 100 prompts from PMD with the original filters and random seed: 20 for each of the five subject types (portrait, landscape, character, object, abstract). The sample therefore covers the same range of subjects as before at four times the size, and the original 25 prompts are a subset of the new 100. We also enlarge the low-prevalence categories (Repeating 10 → 24, Magic 10 → 30) and keep the same two models (FLUX.1-schnell, SDXL 1.0) and 4 images per variant, for 7,880 new images.

Three quantities: correction, effect sizes and steering rates. We keep the original decision rule: a category is steering if it shows image-level influence (SSIM and LPIPS both significant) on at least one model and a positive MSCS. Next to each p-value we add:

  • Holm-corrected p. RQ1 tests eight categories per model and metric. At p < 0.05 per test, the chance that at least one category passes by luck is about 34%. The Holm–Bonferroni procedure adjusts the eight p-values to keep this chance below 5%.
  • Rank-biserial r. This is the standard effect size for the Wilcoxon signed-rank test. It ranges from −1 to 1 and measures how consistently adding category c moves the image toward the full-prompt output. Unlike p, it does not grow with sample size, so it separates a strong effect from one that only crosses p < 0.05.
  • Steering rate. We resample the prompts with replacement 2,000 times and re-apply the image-level criterion each time. The steering rate is the share of resamples in which the category passes. It estimates whether the verdict holds under a different draw of prompts, so it directly quantifies how borderline a decision is.

The following table lists our new results. Table R1. RQ1 at n = 25 (Table 1) and n = 100. SSIM p: two-sided Wilcoxon, n = 25 → n = 100, with the Holm-corrected n = 100 value in parentheses. r and steering rate are for n = 100.

CategorySDXL: SSIM pSDXL: r / steering rateFLUX: SSIM pFLUX: r / steering rateSteering (n = 25 → n = 100)
Trending<.001 → <.001 (<.001)0.90 / 1.00.006 → <.001 (<.001)0.65 / 1.00✓ → ✓ strong
Medium<.001 → <.001 (<.001)0.74 / 1.00.002 → <.001 (<.001)0.65 / 1.00✓ → ✓ strong
Artist.039 → <.001 (<.001)0.63 / 1.00.396 → <.001 (<.001)0.48 / 0.99✓ → ✓ strong
Atmosphere.004 → <.001 (<.001)0.49 / 0.99.024 → .072 (.231)0.21 / 0.42✓ → ✓ borderline
Movement.853 → .027 (.106)0.26 / 0.60.005 → .003 (.014)0.34 / 0.85✓ → ✓ borderline
Quality.134 → .028 (.106)0.25 / 0.60.958 → .058 (.231)−0.22 / 0.00× → ×
Repeating.275 → .087 (.173)−0.40 / 0.00.049† → .391 (.391)0.20 / 0.09× → ×
Magic.232 → .670 (.670)0.09 / 0.07.695 → .119 (.238)−0.33 / 0.00× → ×

†: significant in the wrong direction. SSIM is the binding criterion: LPIPS and MSCS are significant after correction for all six jointly sampled categories on both models. Repeating and Magic use n = 24 and n = 30. In the enlarged run each image seed is shared across the variants of a prompt, so we compare p-values and verdicts rather than absolute SSIM values.

Quality remains a non-steering category. On FLUX, adding Quality moves images away from the full-prompt output: SSIM decreases (r = −0.22) and the steering rate is 0.00. On SDXL, Quality does not pass either after correction (p = .106). It therefore shows image-level influence on neither model. Although quality may express its own semantics, it does not move the image structure toward the full-prompt output, which is the part of the definition it fails. Both strong and borderline categories count as steering under our definition (steering on at least one model), so the five categories are the same as in the original paper.

Comparison with Quality category. Please see below in Response to Reviewer QGVG (Part 2).

AuthorsOpenReview ↗

Response to Reviewer QGVG (Part 2)

Comparison with Quality. Quality also behaves as a non-steering category in RQ2. We generate the top-5 PMD Quality terms (concept art, highly detailed, sharp focus, intricate, smooth) with the identical RQ2 protocol on FLUX, SD3.5 and Midjourney V7 (900 images).

  • Bias. Under the VLM annotator, which agrees with our human labels on 99% (gender) and 97% (race) of images, Quality produces the smallest mean gender shift (|ΔGBS| = 0.11 vs. 0.13–0.25 for the steering categories). It also produces the fewest significant gender shifts (8% of cells vs. 17–35%). The original ViT annotator gives the same result (|ΔGBS| = 0.12 vs. 0.14–0.25).
  • Deepfake detectability. For each category, we test whether the detection rate differs from the baseline in 12 cells (3 models × 2 subjects × 2 detectors. Quality changes the detection rate significantly in only 1 of 12 cells, the fewest of all categories.
  • Safety. With benign PMD spans, Quality changes the NSFW rate least (0.037 vs. 0.063–0.087), about half the effect of any steering category.

Together, these results show that the steering categories selected in RQ1 drive larger downstream changes than the non-steering categories, which supports that our RQ1 selection matters.

W2. Reliability of the FairFace classifiers.

The 100-image validation in the paper (Section 5, Human Validation) uses a stratified sample: 5 images from each of the 6 categories (baseline and the five steering categories) for each of the 3 models (90 images), plus 10 randomly drawn images to reach 100. We agree that 100 images is a small sample, especially for estimating accuracy per category. We therefore draw a second stratified sample of 360 images (20 per model × 3 models × 6 categories), with no overlap with the original 100, giving 460 images in total.

Human annotation results. Two authors independently label each image for the presence of a person, perceived gender and perceived race. Agreement is computed on the 448 images that both annotators mark as containing a person. We obtain the following results.

SampleGender agreementGender κRace agreementRace κ
Original 10095/100 (95.0%)0.9191/100 (91.0%)0.81
New 360338/348 (97.1%)0.95321/348 (92.2%)0.83
Total433/448 (96.7%)0.94412/448 (92.0%)0.82

The two annotators then resolve every disagreement by discussion and agree on a single consensus label. We compare the two FairFace classifiers used in the paper, FairFace Gender Image Detection for gender and fairface-race-vit for race, against these consensus labels, separately for each category. We have the following results.

ClassifierBaselineArtistAtmosphereMediumMovementTrending
Gender (FairFace Gender Image Detection)0.950.920.910.970.880.93
Race (fairface-race-vit)0.920.820.910.840.860.92

FairFace classifiers reliability. Gender accuracy stays between 0.88 and 0.97 in every category. The lowest value is Movement (0.88), whose styles depart furthest from photographs, and it remains close to the photographic baseline (0.95). Race accuracy is 0.92 at baseline and stays between 0.82 and 0.92 with modifiers. Trending matches the baseline (0.92) and Atmosphere is almost unchanged (0.91), while Movement (0.86), Medium (0.84) and Artist (0.82) are lower. Artistic modifiers therefore lower race accuracy only moderately, by at most 0.10, and every category stays above 0.80.

Therefore, these results show that the FairFace classifiers are reliable for gender and mostly reliable for race on images generated with artistic modifiers. In the revision, we will report the 460-image validation and also include the per-category accuracy table in our revised version.

AuthorsOpenReview ↗

Response to Reviewer QGVG (Part 3)

W3. Direction of demographic shifts and their interpretation as social risk.

Demographic shifts is deliberately designed to be non-directional. As stated in Section 4.2, it evaluates whether steering modifiers shift demographic representation relative to the subject-only baseline, which serves as the control. A modifier category can affect bias in either direction, by increasing or by reducing demographic imbalance, and both count as an effect of the modifier. We therefore use a chi-square test with Cramér’s V, which measures how far a distribution moves from the baseline regardless of direction. Whenever GBS or RBS under a modifier differs significantly from the baseline, we conclude that the modifier category affects bias.

We, however, agree with the reviewer that reporting the direction makes this effect easier to interpret, so we separate the two directions. For gender, a shift increases imbalance when GBS rises and reduces it when GBS falls. For race, a shift increases imbalance when RBS falls (the racial composition becomes more concentrated) and reduces it when RBS rises. We obtain the following results.

CategoryMean GBS (baseline 0.86)Sig. gender shifts: reduce / increase imbalanceMean RBS (baseline 0.48)Sig. race shifts: reduce / increase imbalance
Artist0.8016 / 20.398 / 4
Atmosphere0.7613 / 20.4714 / 7
Medium0.824 / 10.398 / 6
Movement0.7011 / 20.374 / 9
Trending0.817 / 00.3610 / 3
Total0.7851 / 70.4044 / 29
  • Gender. Every steering category lowers the mean GBS relative to the baseline (0.70–0.82 vs. 0.86), and most significant shifts (51 of 60) reduce gender imbalance. All 7 shifts that increase it occur for teacher, where a baseline of about 84% women becomes 90–100% men, e.g., “a teacher, greg rutkowski” and “a teacher, dramatic lighting” on FLUX.1-dev.
  • Race. The average direction is the opposite. Four of the five categories lower the mean RBS (0.36–0.39 vs. 0.48), and Atmosphere leaves it nearly unchanged (0.47). Among significant shifts, 44 reduce and 29 increase imbalance. For instance, for engineer, nurse and teacher, increases outnumber reductions (29 vs. 16). Interestingly, in 8 cases, the generated set collapses to a single racial group (RBS = 0), e.g., “a nurse, anime” on FLUX.1-dev.

Both directions are a social risk. For gender, the fair reference is parity (50% men and 50% women, GBS = 0). The baseline already shows that the models themselves are biased: its mean GBS is 0.86, and in 11 of 12 baseline cells at least 75% of the generated people share one gender. Modifiers move some of these distributions toward parity, but they rarely reach it. Under modifiers, 80% of cells still have at least 75% of one gender, and only 12% fall within 40–60%. Even the most dramatic shifts do not remove the imbalance. E.g., on FLUX.1-dev, “a CEO, alphonse mucha” and “a CEO, artgerm” turn 100% men into 100% women, so the gender bias remains at its maximum (GBS = 1).

For race, the most balanced composition over our four groups gives RBS = 0.75, while modifier outputs average 0.40. Hence, whether a modifier increases or reduces imbalance relative to the baseline, the resulting distribution remains far from parity, and the modifier changes the demographic makeup in a way the user does not ask for. Therefore our design counts any significant shift as a social risk. In the revision, we will report its direction together with its size, and state the direction of each category’s effect.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision of Submission #60929 Submitted

Dear Area Chair and Reviewers,

Thank you for all your comments and feedback!

We have uploaded the revised version of our paper. All changes are marked in blue.

We have tried best to address all comments from the reviewers and follow-up discussions, and revised the paper accordingly. The revision also reports the key new experimental results in the main text and the appendix.

Thank you again for your time and constructive feedback!

Best regards,

Authors of Submission #60929

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.