Causally Motivated Diffusion Sampling Frameworks for Harnessing Contextual Bias

ICLR 2025 Conference Submission1637 Authors

Keywords: Causal Inference, Diffusion Models, Contextual bias, Spurious Correlations, Object Cooccurrence, StableDiffusion

Abstract: Diffusion models have shown remarkable performance in text-guided image generation when trained on large-scale datasets, usually collected from the Internet. These large-scale datasets have contextual biases (e.g., co-occurrence of objects) which will naturally cascade into the diffusion model. For example, given a text prompt of ``a photo of the living room'', diffusion models frequently generate a couch, a rug, and a lamp together while rarely generating objects that do not commonly occur in a living room. Intuitively, contextual bias can be helpful because it naturally draws the scene even without detailed information (i.e., visual autofill). On the other hand, contextual bias can limit the diversity of generated images (e.g., diverse object combinations) to focus on common image compositions. To have the best of both worlds, we argue that contextual bias needs to be strengthened or weakened depending on the situation. Previous causally-motivated studies have tried to deal with such issues by analyzing confounders (i.e., contextual bias) and augmenting training data or designing their models to directly learn the interventional distribution. However, due to the large-scale nature of these models, obtaining and analyzing the data or training the huge model from scratch is beyond reach in practice. To tackle this problem, we propose two novel frameworks for strengthening or weakening the contextual bias of pretrained diffusion models without training any parameters or accessing training data. Briefly, we first propose causal graphs to explicitly model contextual bias in the generation process. We then sample the hidden confounder due to contextual bias by sampling from a chain of pretrained large-scale models. Finally, we use samples from the confounder to strengthen or weaken the contextual bias based on methods from causal inference. Experiment results show that our proposed methods are effective in generating more realistic and diverse images than the regular sampling method.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1637