acceptodds
Under review as a conference paper at ICLR 2027

Just as Fair: Balancing Diffusion Trajectory for Fair text-to-Image Generation

Abstract

Diffusion models exhibit persistent demographic biases, systematically favoring some attributes over others in text-to-image generation. While prior work largely address bias in prompts or representations, we ask a different question: does the denoising process itself explore demographic attributes unevenly? We find that it does, biased generation is closely associated with an imbalance in the exploration of denoising trajectories, which reflects statistical biases in the training data and persists throughout sampling. This motivates a trajectory-level view of fairness, where fair generation is treated as balanced exploration of demographic attribute modes, rather than solely as a property of the final outputs. Based on this view, we develop a trajectory-level debiasing framework for fair text-to-image generation. We first introduce trajectory bias, together with an observation mechanism based on Tweedie's formula, which estimates the semantic outcome of intermediate noisy latents and makes attribute preferences measurable throughout denoising. Building on this signal, we dynamically regulate the denoising trajectories with a balanced allocation budget, steering the sampling process toward fair attribute coverage. A key advantage of our approach is its architecture-agnostic nature, which intervenes directly in sampling dynamics and overcomes a core bottleneck of existing debiasing methods. Across four representative diffusion architectures, including U-Net, DiT, and rectified flow models, our method consistently improves demographic balance while preserving generation quality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fairness, Diffusion Model, Text-to-Image Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2Mk7OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper solved the fairness problem from the denoising process. The authors developed a trajectory level debiasing framework for fair text to image generation. More specifically, they formalized the trajectory level fairness as a balanced assignment problem which compute a soft assignment matrix over different attributes. This problem then was solved using Sinkhorn algorithm. The experiments verified the effectiveness of the proposed method in improving demographic balance while preserving generation quality compared to four representative diffusion architectures. The authors listed the limits of this method in time complexity and memory consumption and proposed it as a future work.

Critical strengths

This paper is well written and easy to follow. The authors presented the idea well. The background introduction is complete, the math part (the optimal transport problem, the Tweedie’s denoising estimator are well-known) is correct, and the experiments and ablations studies are organized very well to show the advantages of the proposed method in improving fairness while preserving prompt semantics and image quality compared to existing baselines and the limits of this method in time complexity.

Critical weaknesses

The main weakness, as the authors pointed it out already, is that this method increases the inference time by approximately 6 times and peak memory usage by approximately 1.5 times relative to base generation.

There are also some minor issues: 1.The link can not be opened: https://anonymous.4open.science/ r/Just_As_Fair/. 2.It should be a full stop . instead of , in equation 3 since this sentence ends there. 3. It should be a , in Equation 7. 4. In proposition 1, “thus the balanced exploration loss provides a upper...”---> an upper 5.Under the table 5, there is a sentence “and training lightweight estimators to for more...”, I think you mean “estimators for more...” in this sentence.

Questions for authors

1.The proposed method looks quite similar to the idea in Balanc. Act, and In Table 2, it seems that Balanc. Act has a better fairness value than the proposed method in the Race column, can you explain why. 2.Can you compare Balanc. Act with the proposed method in other DM frameworks except SD v1.5 and attributes except the gender and race. The authors may need more experiments and comparison between these two to convince the readers that the proposed method is better than the other one except saying in the paper: “ while the best performing baseline, Balancing Act, requires substantial additional training at the cost of prompt alignment.

Reviewer_6HzvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes debiasing text-to-image diffusion models by balancing the sampling trajectories across a batch of samples. The main idea behind the method is to exploit the mean-estimate of final denoised sample using Tweedie's trick to obtain a "fairness score" using an attribute classifier. Specifically, at a given denoising step, the authors estimate the clean latent x^0θ(xt,t)\hat{x}_0^\theta(x_t, t) using the noisy latents xtx_t. Then this estimate is passed through an attribute classifier to measure the attribute scores. During inference, inside a slected time interval, the authors propose to solve an entropy-regularized OT assignment problem across a batch of NN parallel trajectories to compute a soft target assignment matrix QtQ_t that satisfies uniform marginal column constraints while staying close to each trajectory's current attribute tendency. The latents are then updated in the gradient direction of the OT objective. The authors show experiments across differerent architecture spanning stable diffusion (SD) v1.5, v3, PixArt-Σ\Sigma, and FLUX.1.

Critical strengths
  1. The "architecture-agnostic" nature of the method is appealing. Usually, inference-time debiasing methods rely on a good semantic space (like H-space of UNet-based architecture). However, the proposed method and experiments show that their method works as well without such constraints.
  2. Formulating debiasing as dynamic batch-assignment via OT is well motivated. While this is well formalized, I don't think this is new. Some previous work already do this, although not explicitly formulated as OT (see weaknesses).
  3. The experiments are comprehensive spanning different T2I backbones and attributes. The medical case study is appreciated, such usecase often gets undermined.
Critical weaknesses
  1. The novelty of proposed method is overstated. Specifically, inference-time distribution guidance along sampling trajectory, Tweedie lookahead probing, and dynamic batch-level assignment to avoid forcing samples across decision boundaries are already explored extensively in BalancingAct [1] and Score Guidance [2].
  2. In Eq. 12, if we substitute back the definition H(⋅)\mathcal{H}(\cdot), the objective becomes minimization KL divergence, DKL(Q∣∣P)D_{KL}(Q||P) and the constraint ensures that QQ is a valid distribution. This is exactly what BalancingAct optimizes for. I think the only difference between the two methods is the way this objective is optimized.
  3. Score Guidance is a critical work missing from the literature and comparisons. While the proposed method requires an attribute classifier for each attribute, Score Guidance operates in the same setting without any such requirement. There are other similarities/conceptual overlaps between the two works like the way both handle multi-attribute setting by using cartesian product of classes (current work) and recursive tagging (in score guidance). Given these pointers, I think the authors should clarify the position of current work w.r.t Score Guidance.
  4. The computational overhead is prohibitively high. In Table 5, one can see about 5×5\times to 6×6\times increase in inference time. In modern workloads, inference occupies a significant proportion. I understand the trade-off since the proposed method is primarily an inference-time intervention, however, 6×6\times increase is too much. Further, I would request the authors to plot this time against number of trajectories, NN. Also, at what NN does the method go out of memory (OOM)?
  5. Proposition 1 is a very trivial application of Pinsker's inequality and adds very little theoretical depth into actual dynamics of underlying SDE/ODE.

[1] Parihar, Rishubh, et al. "Balancing act: Distribution-guided debiasing in diffusion models." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.
[2] Tiwary, Piyush, and Prabhav Verma. "Debiasing Diffusion Models via Score Guidance." Transactions on Machine Learning Research.

Questions for authors
  1. By the OT formulation, is it necessary that N≥CN\ge C? If yes, then how does the method work when N<CN< C?
  2. In Eq. 14 (dti=∇xtiLBE(Xt)d_t^i = \nabla_{x_t^i} \mathcal{L}_{BE}(X_t)), since x^0θ(xt,t)\hat{x}_0^\theta(x_t, t) is a function of ϵθ(xt,t)\epsilon_\theta(x_t, t) (Eq. 6), do you backpropagate through the denoiser ϵθ(xt,t)\epsilon_\theta(x_t, t) when computing ∇xtiLBE\nabla_{x_t^i} \mathcal{L}_{BE}, or do you apply a stop-gradient to ϵθ(xt,t)\epsilon_\theta(x_t, t)? I am also slightly surprised as in how does FLUX.1 with a batch size of N=16N=16 fit into 24.26 GB of VRAM on an RTX 4090 (Table 5)?
  3. How does the proposed method compare theoretically and empirically with Score Guidance (Tiwary et al., TMLR 2026), which already established the theoretical link between unconditional/marginal scores and mode dominance along diffusion trajectories, and utilized Tweedie x^0∣t\hat{x}_{0\vert{}t} estimates for training-free batch-level trajectory regulation without requiring external supervised classifiers?
Reviewer_TXekOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an inference-time method for observing demographic attributes at intermediate denoising states of text-to-image diffusion models and balancing the attribute distribution within a batch. At selected intervention steps, the method decodes Tweedie-based clean-image estimates to obtain classifier probabilities and computes sample-level assignments using Sinkhorn, subject to equal total allocation across attributes. It then intervenes in the sampling trajectories by differentiating a cross-entropy loss with respect to the latents, treating the assignments as stop-gradient targets.

The method is evaluated on SD v1.5, SD 3, PixArt-Σ, and FLUX.1, reporting improvements in final gender and race fairness metrics. The appendix also includes intersectional attributes, non-human prompts, classifier and intervention-window variations, and a dermatological image generation case study. However, the theoretical guarantee is an upper bound on the imbalance of mean classifier probabilities. How tightly this bound relates to actual group coverage or improvements in the final sampling outcomes remains unclear.

Critical strengths
  • The paper demonstrates a training-free control method that can be applied across multiple diffusion architectures. Rather than relying on specific hidden representations within the model, it uses intermediate latents as a common intervention point. The method is evaluated on four models (SD v1.5, SD 3, PixArt-Σ, and FLUX.1) and the gender and race results in Table 1 show improvements over the base models in both Fairness Discrepancy (FD) and Fairness Score. Its applicability across different architectures without additional generator training is a practical advantage.

  • The design of distributing a batch-level objective through sample-level assignments is concrete, and the evaluation covers a relatively broad range of settings. The Sinkhorn assignment, which allocates equal total mass to each attribute while accounting for the current classifier scores, is a reasonable design choice. By considering each sample's current tendency, it may achieve the desired balance with smaller modifications than enforcing predetermined hard labels. Although this potential advantage requires a direct comparison, the motivation behind the method is clear.

Critical weaknesses
  • There is an important gap between the surrogate being controlled and the fairness property the paper aims to achieve. Definition 3.1 defines the trajectory attribute score as a logit, whereas the subsequent loss and Proposition 1 assume probabilities. The conversion to probabilities, including softmax and temperature calibration, should be clarified. Even when interpreted as probabilities, balancing the mean classifier scores is not equivalent to achieving balanced hard group coverage. For example, if 90% of samples have a class-1 probability of 0.51 and the remaining 10% have a probability of 0.41, the mean probability is exactly 0.5, but the argmax-based group ratio is 90:10. Therefore, changes that increase classifier uncertainty should be distinguished from changes that achieve actual demographic allocation.

    Furthermore, Proposition 1 establishes the following upper bound at a particular state: B(t)≤2C(LBE−1N∑iH(qi)).B(t)\leq\sqrt{\frac{2}{C}\left(L_{\mathrm{BE}}-\frac{1}{N}\sum_i H(q_i)\right)}.

    This does not establish that a finite update in Eq. (16) always decreases B(t)B(t), that the subsequent denoising transition preserves such a decrease, or that the final generated images have balanced hard-category assignments.

    Appendix C.3 already reports a group-count-based Fairness Score, using an evaluation classifier different from the guidance classifier. My concern is therefore the scope of the theoretical guarantee and the relationship between the intermediate surrogate and final group allocation, rather than an absence of final-outcome evaluation.

  • The additional contribution of dynamic assignment and its effectiveness relative to computational cost are not sufficiently isolated. Although the paper compares against several existing methods, it does not include comparisons with a direct penalty that matches the batch-mean classifier probabilities to a uniform distribution, a fixed balanced assignment, or an assignment computed only once, under the same classifier and intervention window. The intervention-window, prompt, and classifier ablations in Appendix E.4 do not replace these controls. Therefore, it remains unclear how much of the reported improvement can be specifically attributed to dynamic Sinkhorn assignment.

    Computational cost is also an important consideration. According to Table 5, the inference-time overhead relative to the base models is 5.86× for SD v1.5, 5.01× for SD 3, 5.91× for PixArt-Σ, and 4.72× for FLUX.1, while peak memory usage increases by 1.28–2.29×. To better establish the practical cost-effectiveness of this training-free approach, comparisons with attribute-conditioned generation or generate-and-select methods under a matched inference-time budget would be useful.

Questions for authors
  • Definition 3.1 describes the attribute score as a logit, whereas Eq. (11) and Proposition 1 require probabilities. What exact probability transformation and temperature, if any, are used for guidance? Since Proposition 1 bounds soft-probability imbalance rather than guaranteeing balanced final group counts, could the authors quantify how intervention-induced changes in B(t)B(t) relate to the final group-count-based Fairness Score, while also monitoring classifier confidence or entropy?

  • How much of the improvement specifically depends on recomputing the balanced Sinkhorn assignment? Under the same classifier, batch size, and intervention window, could the authors compare against a direct batch-mean probability penalty and a fixed or one-time balanced assignment, reporting final fairness, CLIP-T/MUSIQ, and runtime? Given the reported 4.72–5.91× inference time, a comparison with a simple generate-and-select baseline at the same wall-clock budget and number of retained images would also help establish the practical benefit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.