Test-Time Modification: Inverse Domain Transformation for Robust Perception

2025-12-15

AI Summary

For domain generalization under unknown target domains, this paper proposes the Test-time Inverse Domain Transformation (TIDT) paradigm: leveraging diffusion models to inversely map target-domain images in latent space back to the source-domain distribution, enabling pre-trained downstream models to generalize robustly without fine-tuning. The method requires only source-domain text descriptions—no large-scale synthetic data generation—and is compatible with black-box task models and multi-source diffusion models. Core technical contributions include diffusion latent-space inverse transformation, source-domain semantic-guided reconstruction, and multi-model ensemble for robustness enhancement. Evaluated on BDD100K-Night (segmentation), ImageNet-R (classification), and DarkZurich (detection), TIDT achieves relative performance gains of 137%, 68%, and 62%, respectively, unifying support across three major vision tasks while significantly reducing deployment costs for domain adaptation.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• User Modeling, Personalization and Recommendation: User modeling for targeted and personalized online advertising
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Generative foundation models contain broad visual knowledge and can produce diverse image variations, making them particularly promising for advancing domain generalization tasks. While they can be used for training data augmentation, synthesizing comprehensive target-domain variations remains slow, expensive, and incomplete. We propose an alternative: using diffusion models at test time to map target images back to the source distribution where the downstream model was trained. This approach requires only a source domain description, preserves the task model, and eliminates large-scale synthetic data generation. We demonstrate consistent improvements across segmentation, detection, and classification tasks under challenging environmental shifts in real-to-real domain generalization scenarios with unknown target distributions. Our analysis spans multiple generative and downstream models, including an ensemble variant for enhanced robustness. The method achieves substantial relative gains: 137% on BDD100K-Night, 68% on ImageNet-R, and 62% on DarkZurich.

Problem

• Enhancing domain generalization via test-time inverse domain transformation.
• Eliminating need for large-scale synthetic target-domain data generation.
• Improving model robustness under unknown environmental distribution shifts.

Innovation

• Test-time diffusion models transform target images
• Maps target images back to source distribution
• Eliminates large-scale synthetic data generation