Debiasing Text-to-image Diffusion Models with Self-discovering Latent Directions

1. (Major) Novelty:

In the model, it seems like the main novelty comes from the self-discovering training process of the adapters for each attribute category. However [1] already explores the idea of creating category directions for attributes. [1] discovers attribute directions by fine-tuning LoRA adaptors on a diffusion model with the loss: $\epsilon L = \epsilon\theta(X, c_t, t) + \eta(\epsilon\theta(X, c_+, t) - \epsilon\theta(X, c_-, t))$

If the negative prompt $c_-$ is replaced by a neutral prompt $c_t$, it becomes the identical to the loss used in this paper. So, in terms of the loss formulation, it is similar to [1]. From an architectural perspective, while [1] uses LoRA, this work uses 1-D adapters (as used in [2]), different approaches both aiming to alleviate full fine-tuning of the DM. The modification from [2] seems to be that it is deploying the adapters in only the cross-attention layers rather than the whole Unet, which as a stand-alone does not amount to much novelty.

2. Intersectional Debiasing:

The paper employs an orthogonal loss between attributes when tackling multi-attribute debiasing. This dependency during training hampers the reusability aspect of these adapter. Ex: If I were to individually debias gender, race and age, I would train adapter directions for each. But now if I want to debias 2 of them, or all 3 attributes at a time, I cannot reuse these learnt single attribute direction vectors. This problem becomes more apparent especially in the case of intersection of multiple attributes.

Although I do understand that the orthogonality loss during training improves the numbers (as shown in the ablations section), it would be interesting to explore how these models trained for individual attribute directions can be leveraged for multi-attribute debiasing, without needing to retrain them for each subset of attributes.

3. (Minor) Architectural Diagram: The equation in Figure 4 has an error that needs correction:

$\epsilon\theta(X_t, g_t, t) + \eta(\epsilon\theta(X_t,$ d$, t) - \epsilon\theta(X_t, g_t, t))$

References:

[1] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, David Bau. "Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models" arXiv preprint arXiv:2311.12092 (2023).

[2] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R.Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models, 2024. https://arxiv.org/abs/2402.18206

[3] Mengyao Lyu,Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin,Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications, 2024. https://arxiv.org/abs/2312.16145.

• The results are shown on limited types of biases (only gender and race). Evidence that the method for other kinds of biases might be helpful (see questions).

• All results are shown single person or single occupation. What if multiple people are generated for a single occupation?

• All results are shown for a fixed ratio of bias (i.e, generations are balanced across the chosen bias attribute). Ablations should be there for different ratios as shown in ITI-Gen [a]

• Comparisons with competitive methods are less (e.g.,[a, b]).

    [a] Zhang, Cheng, et al. "Iti-gen: Inclusive text-to-image generation." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.
   [b] Li, Hang, et al. "Self-discovering interpretable diffusion latent directions for responsible text-to-image generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.
  

1. The optimization process for learning adapters closely resembles [1], potentially limiting the method’s novelty.
2. The study omits comparisons with prior self-discovering debiasing methods, such as [2], which learns a latent vector for each concept. This further limits the uniqueness of the contribution.
3. Comparison protocols differ from previous works ([1], [4]), omitting the Winobias dataset ([3]). Justifying this deviation would clarify the appropriateness of the comparisons.

[1] Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models (ECCV 2024). [2] Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation (CVPR 2024). [3] Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods (NAACL 2018). [4] Unified Concept Editing in Diffusion Models (WACV 2024). [5] Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2023).