acceptodds
Under review as a conference paper at ICLR 2027

The Prompt Effect: How Conditioning Strength Shapes Stereotype Differentiation in Text-to-Image Diffusion Models

Abstract

Bias audits of text-to-image diffusion models typically evaluate generations produced under explicit semantic prompts, even though measured demographic differentiation may itself depend on how strongly those prompts condition generation. We study this by comparing an operational weak-conditioning baseline, empty-prompt generation at classifier-free guidance (CFG) scale 1, against neutral prompted generation across six gender–occupation and social contexts. On Stable Diffusion v1.5, prompted generations show consistently greater per-image stereotype differentiation than the weak-conditioning baseline (LBAS < 1 in all six contexts), robust to relevance filtering and normalization. We then intervene on conditioning strength, holding prompt and seed fixed. Increasing CFG from 1.0 to 7.5 increases |skew| by 0.00747 on average across 36 matched context–seed pairs (two-sided Wilcoxon p = 0.00136), and a context fixed-effects regression estimates a positive CFG effect (β = 0.001079, p = 0.00121). However, CFG also increases semantic relevance, and the CFG–skew association attenuates to approximately zero after adjusting for relevance, suggesting the amplification is closely coupled to stronger semantic alignment rather than a relevance-independent direct effect. The directional pattern also replicates on SDXL and under OpenCLIP rescoring. These results characterize stereotype differentiation as a property of the conditioning pipeline, not just the underlying model, and motivate fairness audits that evaluate behavior across conditioning strength rather than at one prompted operating point.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion models, text-to-image generation, fairness, social bias, stereotype amplification, classifier-free guidance, generative AI

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_4tetOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates how cueing and CFG strength affect gender-related measures of text-to-image models. The authors compare an empty cue, a weakly conditioned baseline of CFG=1 with six categories of neutral cues, and fix the cue and initial seed scan CFG. Based on the absolute score difference of the CLIP control text, the paper reports that the cued images are more distinguishable, and the increase in CFG is also accompanied by an increase in discrimination; the correlation is close to zero after controlling for semantic relatedness. The SDXL and OpenCLIP experiments replicate the general direction.

Critical strengths

A rigorous ablation study was conducted. Including:

  1. A CFG intervention that fixes hints and seeds is more explanatory than an audit of a single default configuration and makes a clear distinction between controlling initialization and leaving the final image unchanged (§3.1).
  2. The authors are candid in stating that LBAS does not measure net directionality, relevance is a post-intervention variable, and thus the adjusted results do not identify independent direct causal effects (§§3.3, 3.6, 4.5).
Critical weaknesses

This paper suffers from significant readability issues. First, the introduction is too brief and fails to adequately present the motivation behind the research. Furthermore, the list of contributions lacks consistent formatting, and Figure 1—a key visual—would be better conveyed through text rather than as an image.

Second, the impact of a prompt is highly dependent on the specific model employed, making such findings time-sensitive. The fact that this paper relies on a U-Net-based architecture—without offering any architectural insights—is a critical flaw.

Finally, the "Limitations" and "Future Work" section contain a substantial amount of material that should actually fall within the scope of the paper’s core research. This indicates that the manuscript requires further refinement.

Questions for authors

See weakness.

Reviewer_AYh5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

To understand societal biases in diffusion-based text-to-image models, it is common to analyze images generated using classifier-free guidance (CFG). However, this analysis conflates the contributions of the models' inherent biases with biases amplified by text conditioning and CFG. Instead, the authors propose measuring biases in images generated with CFG and in unconditional images generated from null prompts. They propose metrics called skew and LBAS to measure the effect of text conditioning on gender biases in generated images. Furthermore, they examine the relationship between different CFG scales and the skew of generated images. Experiments using Stable Diffusion 1.5 and SDXL as backbone models and CLIP (ViT-B/32) and OpenCLIP (ViT-L/14; LAION 2B) as evaluators support their claims.

Critical strengths

The authors attempt to analyze societal biases in diffusion models by comparing unconditional images generated from null prompts with text-conditioned images generated using CFG. This approach could help disentangle the inherent biases learned by the models from biases introduced through CFG.

Critical weaknesses
  1. Unclear motivation and contribution
    1. In the abstract, the proposed metrics, LBAS and skew, are introduced without explanation. Readers cannot understand their meanings or scales. For example, it is unclear how to interpret “LBAS < 1 in all six contexts” (L20-L21) or “increasing CFG from 1.0 to 7.5 increases |skew| by 0.00747 on average” (L22-L23). Since proposing LBAS is one of the contributions of this work, the authors should explain the purpose and interpretation of this metric in the abstract.
    2. In the introduction, the authors point out that conventional prompt-based fairness auditing methods interpret demographic skews observed under text-conditioned CFG as model-intrinsic biases (L40-L43). However, they cite no prior work to support this claim, and it is unclear how existing work has adopted this interpretation. Moreover, the introduction does not clearly articulate which problems remain unsolved in fair image generation or how this work addresses them. In particular, the scientific and practical necessity of comparing null-prompt and text-conditioned generations cannot be justified solely by the fact that no prior work has investigated such comparisons.
    3. The authors present LBAS and an analysis of CFG scales as their contributions (L54-L62), but the novelty of these contributions is not clearly established. The effects of CFG scales on biases in generated images have already been investigated by Kim et al. (2026). The authors should clarify how this work provides new insights beyond prior work. Furthermore, although LBAS is proposed as a novel metric, the main CFG-scale analysis (Sec. 4.5) primarily relies on absolute skew, and the role of LBAS itself remains limited. In particular, the paper does not demonstrate what additional information LBAS provides beyond conventional metrics.
  2. Validity of the proposed metric and measurement
    1. The authors also propose filtered LBAS to compensate for domain inconsistencies between unconditional generation and conditional generation using occupational prompts by selecting the top 50% of null-prompt images ranked by relevance to each context (Section 3.4). However, it remains unclear whether this filtering procedure retains enough images to enable meaningful evaluation of gender skew. Based on my visual inspection of the 16 dream samples shown in Fig. 2, only two images appear to contain clearly identifiable people (row 2, column 2, and row 4, column 3). Furthermore, I can assess gender presentation in only one image (row 4, column 3). I am concerned that retaining only the top 50% of images by relevance may not be sufficient to select images that meaningfully depict people and gender. The authors should report the proportion of filtered images that contain people, their relevance to the target occupations, and the extent to which gender presentation can be meaningfully assessed to address this concern.
    2. A comparison of Tables 2 and 3 reveals that the average absolute skews for random noise are close to those for dream images in some occupational contexts. For instance, in Nursing, the absolute skew for noise (0.004249) is about 65% of that for dream images (0.00654), and in Housework, the absolute skew for noise (0.005149) is approximately 62% of that for dream images (0.00829). I wonder whether the CLIP-based skew score can capture meaningful demographic signals in dream images, given that semantically meaningless noise images can yield skew values comparable to those of dream images. Furthermore, the presentation of the noise control experiment (Sec. 4.3) is defensive and methodologically unconvincing. Rather than establishing why the proposed random-noise baseline is appropriate, the authors describe this control as “conservative but imperfect,” acknowledge its weaknesses, and then propose an ostensibly superior alternative that they neither explain in detail nor implement. If the authors are already aware of a better control, they should explain why they did not use it instead of asserting that an unperformed experiment would be a stronger alternative.
    3. Although the study aims to analyze stereotypes, its experiments focus on binary gender biases across six occupational prompts. Stereotypes involving other demographic attributes, such as race, age, geographic background, and body type, are not evaluated. Furthermore, all probes rely on a binary male–female gender contrast, excluding nonbinary gender representations and biases arising from intersections of multiple social attributes. Therefore, the scope of the evaluation is limited, and caution is needed when interpreting the results as evidence of stereotype amplification in general. Moreover, Stable Diffusion v1.5 and SDXL are investigated as backbone diffusion models. However, given advances in open-source text-to-image models, the authors should consider more recent models such as Stable Diffusion 3.5 (Esser et al., 2024), Flux 2 (Black Forest Labs, 2026), or Qwen-Image (Wu et al., 2025) to better support the claims in the paper. Furthermore, the study uses CLIP ViT-B/32 and OpenCLIP ViT-L/14, both of which belong to the CLIP family of evaluators. However, CLIP models are also known to exhibit inherent societal biases. Since it is impossible to completely eliminate biases in evaluator models, the authors should consider evaluators using different methodologies, such as attribute classifiers like FairFace (Karkkainen and Joo, 2021), larger-scale VLMs such as GPT (OpenAI, 2026), or human annotations.
  • Kim et al. (2026). Stay Fair! Ensuring Group Fairness in Diffusion Models across Guidance Scales. NeurIPS 2026.
  • Esser et al. (2024). Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. ICML 2024.
  • Black Forest Labs (2026). FLUX.
  • Wu et al. (2025). Qwen-Image Technical Report. arXiv.
  • Karkkainen and Joo (2021). FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation. WACV 2021.
  • OpenAI (2026). GPT-6 Astra System Card.
Questions for authors
  • Sec. 3.6 lists several statistical methods, including descriptive statistics, Spearman rank correlation, the Wilcoxon signed-rank test, OLS fixed-effects regression, and mixed-effects models. However, it is not clearly stated why each method was selected, which hypothesis each analysis is intended to test, or how the results of the different analyses relate to one another.
Reviewer_kh9rOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper asks whether text conditioning itself shapes the gender stereotype differentiation measured in text to image diffusion models. The authors compare empty prompt generation at minimal guidance against neutral occupational prompts on Stable Diffusion v1.5 and find that prompting consistently increases per image CLIP skew. A seed paired CFG sweep shows that stronger guidance raises measured skew further, though the effect largely vanishes once semantic relevance is controlled. The pattern replicates on SDXL and with an OpenCLIP judge, and the authors recommend auditing bias across guidance scales rather than at one default setting.

Critical strengths
  • The paper is careful about statistical reporting.
  • It's good that the paper list its own limitations specifically.
Critical weaknesses
  • The CFG result is close to tautological, and the paper’s own analysis confirms this. Classifier free guidance amplifies the conditional signal by construction, so stronger guidance pushing toward the majority gender of an occupational prior is the expected behavior. Once relevance is adjusted the CFG association vanishes, leaving only the claim that more on prompt images score as more gendered. The residual relevance normalized effect is weak and remains exposed to the guidance dependent CLIP geometry problem the authors themselves cite.
  • The positioning overstates novelty. Guidance driven amplification has already been shown, as the authors concede. In addition, [1] establishes the same thesis, that demographic skew is a property of the text conditioning pipeline rather than the weights, by varying system prompts and LLM prompt expansion with direct demographic measurement. That work is not cited, and it both weakens the claim that this comparison has received little study and offers a measurement approach the present paper lacks.
  • The core measure captures CLIP differentiation magnitude, not stereotype direction or actual demographic distribution. A sharper image of a cook raises absolute skew whether the cook is a man or a woman, so the dose response is equally explained by images becoming clearer rather than more stereotyped. The directional metric DBAS is defined but never reported, and no human annotation or gender count is provided.
  • (minor) Figure 1 is hard to read. Consider increasing its width.

[1] Park et al. “Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models.” arXiv preprint arXiv:2512.04981 (2025).

Questions for authors
  • I understand the practical recommendation that fairness audits should be evaluated across guidance scales rather than at a single default operating point. However, is the guidance scale not used as a single fixed value in practice (for example, 7.5 for many Stable Diffusion models)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.