Reliable Image Quality Evaluation and Mitigation of Quality Bias in Generative Models

Additional Reference

Thanks for suggesting a missing reference. Although our paper already includes references related to energy-based guidance in text-to-image models—such as Composing Diffusion Models [1], Self-Guidance [2], and Universal Guidance [3]—the suggested reference [4] is indeed a valuable cornerstone in the energy-based guidance literature. We will include this reference in the revised version of our paper.

[1] Liu, N., Li, S., Du, Y., Torralba, A., & Tenenbaum, J. B. (2022, October). Compositional visual generation with composable diffusion models. In European Conference on Computer Vision (pp. 423-439). Cham: Springer Nature Switzerland.
[2] Epstein, D., Jabri, A., Poole, B., Efros, A., & Holynski, A. (2023). Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36, 16222-16239.
[3] Bansal, A., Chu, H. M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., & Goldstein, T. (2023). Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 843-852).
[4] Zhao, M., Bao, F., Li, C., & Zhu, J. (2022). Egsde: Unpaired image-to-image translation via energy-guided stochastic differential equations. Advances in Neural Information Processing Systems, 35, 3609-3623.

Clarification on Equation 3

Thank you for pointing out the lack of explicit notation in Equation 3. While Groups A and B are introduced as two demographic groups in Section 3.1, the page distance between that section and Equation 3 may cause confusion. To improve clarity, we agree that the definitions of Groups A and B should be stated explicitly near Equation 3.

Specifically, Group A and Group B refer to demographic groups such as male and female. The terms $z_t^A$ and $z_t^B$ represent latent variables derived from the input prompt:
“a photo of a {GENDER} who works as a {PROFESSION}.”

Moreover, Groups A and B can also represent other demographic attributes, such as different races, as demonstrated in our rebuttal to Reviewer h2hy. We expect that this formulation can be extended further to accommodate any desired quality bias to mitigate, depending on the fairness objective.

Extension to Other Generative Models

Yes, the DQA framework can be applied to any generative model such as Flux. Since DQA serves as a reliability measure for evaluation metrics such as FID, it does not rely on the performance of the generative model itself—as long as the model produces images and their quality is evaluated using FID.

Moreover, the DQA-Guidance approach is also model-agnostic and can be applied to any diffusion-based generative models.

Extension of Demographic Groups

Thank you for raising this point. Quality bias is not limited to gender; it also extends to other demographic attributes such as race. In our study, we consider four racial groups: Asian, Black, Caucasian, and Indian. We explore two possible directions for extending DQA-Guidance to handle multi-racial bias:

1. Pairwise Group Comparison
   In this approach, we select two racial groups (e.g., Asian vs. Black) as Group A and Group B and apply DQA-Guidance in the same manner as the gender-bias case. This enables a detailed, pairwise analysis of racial bias and its mitigation.

2. All-at-Once Comparison
   Alternatively, we can modify the DQA-Guidance formulation to consider all racial groups simultaneously by replacing the DQA term in Equation 4 with the average pairwise DQA across all race pairs, as defined below:
   $

   \tilde{\epsilon}_{\theta} (z_t) = \epsilon_{\theta} (z_t) + \sigma_t \nabla_{z_t} \left(\lambda_1 , \text{AvgDQA} + \lambda_2 , D\big(f^(\mathcal{I}_{\text{gen}}), f^(\mathcal{I}_{\text{ref}})\big)\right)
   $

where
$

\text{AvgDQA}(\mathcal{G}) = \frac{1}{\binom{n}{2}} \sum_{1 \leq i < j \leq n} \text{DQA}(G_i, G_j; f^*)
$

This formulation generalizes fairness evaluation across all $n$ racial groups by averaging over all possible group pairs.

For ease of implementation and to enable a more detailed analysis of individual racial quality disparities, we adopt the first approach in our additional experiments.

In our experiments, DQA-Guidance improves both the overall image quality and reduces quality disparity across racial groups, consistent with the results reported for the gender case.

Novelty of the Paper

To the best of our knowledge, this paper is the first to address fairness issues in the evaluation metrics used for generated images.

We distinguish two types of bias:

• (a) Bias in the evaluation metric
• (b) Bias in quality in the generated image

Although (b) has been studied, (a) has not received adequate attention. In this work, we identify bias in a widely used evaluation metric, FID, and propose DQA as a reliability measure to detect and quantify such bias. By isolating evaluation bias through DQA by choosing a reliable image encoder, we can more accurately reveal the quality bias in generative models, (b). Furthermore, we introduce DQA-Guidance, a mitigation strategy that guides generative models toward reducing quality bias in the generated images.

Taken together, our contributions are the first to highlight and address the two-fold fairness challenge in generative modeling.

Regarding Comparison Method

As ours is the first work to identify bias in evaluation metrics and to propose a framework for mitigating quality bias in generative models, we were unable to include direct comparisons with existing methods.

Adjustment in the Controlled Dataset

Please see the rebuttal for Reviewer 4STN.

Theoretical Analysis for Equation 1

The denominator captures the total generation shift, i.e., how far the generated distribution is from the reference distribution across all groups. It serves two purposes.

1. DQA measures how large the inter-group disparity is relative to the overall deviation. If both group-specific shifts are small, then even a small difference between them may be meaningful. Conversely, if the model globally generates low-quality outputs, a larger group disparity might be expected and less concerning.
2. Different generative models, encoders, and data domains may exhibit widely varying absolute distances. Without normalization, a group-level bias in numerator is hard to determined to be negligible or severe. Therefore, the denominator anchors the numerator to this global scale.

Improvement via DQA-Guidance

While Figures 6 and 7 illustrate the improvements in both performance and fairness of image quality, we add a table below for better clarity. This table explicitly demonstrates that an appropriate choice of $\lambda_1$ and $\lambda_2$ leads to substantial improvements in both fairness and overall image quality.

While MMD with DINO serves as our primary quantitative evaluation metric for image quality, here we also report Fréchet Distance (FD) for generated images w/ and w/o DQA-Guidance.

Regarding the qualitative results in Figure 7, while some aspects of visual quality are subjective, the improvements are evident. For example, DQA-Guidance helps remove visual artifacts such as extra limbs (e.g., three hands), enhances image characteristics like color (e.g., from grayscale to natural color), and improves coherence between the prompt and the image (e.g., ensuring the nurse to be male along with a male prompt).

Regarding Algorithm 1

While Algorithm 1 may initially appear complex, its underlying logic is straightforward. It identifies subsets of generated images that most strongly increase or decrease the group-level discrepancy in perceived image quality, as measured by DQA.

These subsets support a downstream diagnostic evaluation. Specifically, we use the fair/unfair subsets in classification as a data augmentation to assess whether the quality discrepancy is practically meaningful. We find that models trained on the unfair subset exhibit larger fairness gaps in downstream classification, while those trained on the fair subset show more equitable performance. This demonstrates that the discrepancy captured by the DQA is predictive of real fairness issues, thereby validating its utility as a diagnostic signal for metric reliability.

Results in Table 1

To show the validty of our experimental results in Table 1, we investigate the confidence interval of experimental results.

Regarding the Vague Content

Thank you for pointing this out. The sentence in Section 5 was intended as an introduction to the extension of DQA as a guidance term for diffusion models. We agree that the current phrasing may be vague and potentially confusing. We will revise this section to more clearly explain how DQA can be extended and applied in the context of guidance for diffusion models.

Adjustment in the Controlled Dataset

The degradations we introduce are well-established in diffusion-based generative models literature.

• Weak Classifier-Free Guidance (CFG)
  In CFG, using weak guidance simulates a scenario where the generated image loses coherence with the prompt.
• Fewer Inference Steps
  As noted in [1], images generated with fewer diffusion steps typically exhibit lower visual quality by leaving more residual noise and artifacts.
• Stronger Initial Noise
  In our method, we use in-painting to construct the controlled dataset. We first generate images from text prompts and then modify them to reflect different attributes to maintain contextual consistency. In in-painting, a stronger noise level preserves more of the original image. As a result, the model struggles to apply the desired attribute modification, leading to poor coherence with the target attribute.
• No Refiner
  The SDXL paper explicitly states that the use of a refiner network improves visual quality, meaning removing the refiner leads to a noticeable degradation in image quality.

Each of these modifications is grounded in existing literature, ensuring that the controlled degradations reflect realistic variations in generation quality. [1] Kim et. al., 2024, Model-Agnostic Human Preference Inversion in Diffusion Models

Hyperparameter Sensitivity

As shown in the figure at the following link:
https://drive.google.com/file/d/1Ot1FkMuPYmb0-6vFZZ5vUHmtcw6xgGAq/view?usp=sharing we investigate the impact of hyperparameter variation in the controlled dataset by adjusting the guidance scale in CFG. The right subfigure shows a clear degradation in the generated images as controlled.

In the main paper, we argue that a lower average DQA across degraded images indicates higher reliability of an evaluation metric. Although this conclusion remains consistent (DINO-RN50 as the most reliable), this additional analysis highlights that robustness to degradation could serve as another criterion for evaluating the reliability of an image encoder.

We will include this analysis in the revised version to better emphasize the desired characteristics of reliablilty for quality evaluation.

Analysis for Experimental Results

Please see the rebuttal for Reviewer yKK1.

Regarding Ablation Study

In Figure 6, the left subfigure shows the performance trend when varying $\lambda_1$ while keeping $\lambda_2$ fixed, and the right subfigure shows the trend when varying $\lambda_2$ with $\lambda_1$ held constant. Therefore, the experimental results in Figure 6 already serve as an ablation study, demonstrating the individual effects of each component in the DQA-Guidance.

Cost for DQA Evaluation

Since DQA is a reliability score for evaluation metrics, its cost is not directly comparable to performance metrics such as FID. However, to clarify the computational requirements: DQA involves three quality evaluations, two for the subgroup qualities (numerator) and one for the overall quality (denominator).
Importantly, DQA needs to be computed only once for each encoder to determine the reliability for evaluation. Once a reliable score is identified, there is no need to recompute DQA during future model evaluations.
On the other hand, MMD with DINO is our primary quantitative evaluation metric. Notably, MMD is reported to be approximately 1,000 times faster than the Fréchet Distance used in FID, making it a more efficient choice for large-scale or repeated evaluations.

Cost for DQA-Guidance

We observe the increase in memory usage. DQA-Guidance utilizes an additional image encoder along with gradient computation for the guidance term. As a result, memory usage increased from 10,124MB to 18,626MB due to the gradient computation.

Potential Limitation

Computational Cost

As shown above, DQA-Guidance introduces additional computational cost. However, since ours is the first work to demonstrate that quality bias-based guidance can steer diffusion models toward fairer outputs, it opens a promising direction for fairness-aware guidance. Future work can further explore cost-efficient implementations of such approaches.

Lack of Real Reference Dataset

Although the controlled dataset contains high-quality images, they may not fully align with human-perceived realism or quality standards. A promising future direction is to develop human-validated reference datasets, where quality judgments are collected through perceptual surveys. This would enhance the validity of DQA as a reliability indicator and offer a more robust benchmark for auditing image evaluation metrics.

Possibility of Overfitting

DQA-Guidance leverages group-specific references. However, if the reference set is narrow or unrepresentative, the model might overfit to a particular visual style, reducing diversity or realism. Constructing more diverse and well-controlled reference dataset would improve the generalizability of DQA-Guidance.