Go to NeurIPS 2025 Conference homepagePreventing Shortcuts in Adapter Training via Providing the Shortcuts
Keywords: text to image generation, adapter, personalized generation, controllable generation, spurrious correlation, shortcut learning
TL;DR: We find that rerouting spurious shortcuts in adapter training enables robust disentanglement for text-to-image generation with adapters.
Abstract: Adapter-based training has emerged as a key mechanism for extending the capabilities of powerful foundation image generators, enabling personalized and stylized text-to-image synthesis. These adapters are typically trained to capture a specific target attribute, such as subject identity, using single-image reconstruction objectives. However, because the input image inevitably contains a mixture of visual factors, adapters are prone to entangle the target attribute with incidental ones, such as pose, expression, and lighting. This spurious correlation problem limits generalization and obstructs the model's ability to adhere to the input text prompt. In this work, we uncover a simple yet effective solution: provide the very shortcuts we wish to eliminate during adapter training. In Shortcut-Rerouted Adapter Training, confounding factors are routed through auxiliary modules, such as ControlNet or LoRA, eliminating the incentive for the adapter to internalize them. The auxiliary modules are then removed during inference. When applied to tasks like facial and full-body identity injection, our approach improves generation quality, diversity, and prompt adherence. These results point to a general design principle in the era of large models: when seeking disentangled representations, the most effective path may be to establish shortcuts for what should NOT be learned.
Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)
Submission Number: 13384
Loading