# Back to the Source: Diffusion-Driven Test-Time Adaptation

**Authors:** Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, Dequan Wang  
**arXiv:** [2207.03442](https://arxiv.org/abs/2207.03442)  
**Date:** 2022-07-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.03442v1/

## Key points

- The paper proposes a novel approach for test-time adaptation that improves model accuracy on shifted target data
- Existing methods update the source model by re-training on each target domain, which is sensitive to data amount and order, and hyperparameters
- The proposed diffusion-driven adaptation method (DDA) updates the target data by projecting all test inputs toward the source domain with a generative diffusion model
- DDA shares its models for classification and generation across all domains, trained on the source domain and fixed during testing
- Input adaptation by DDA is more robust than prior model adaptation approaches across various corruptions, architectures, and data regimes on ImageNet-C benchmark
- DDA succeeds where model adaptation degrades on too little data in small batches, dependent data in non-uniform order, or mixed data with multiple corruptions
- The authors provide an extensive analysis of DDA's performance compared to existing state-of-the-art methods such as fine-tuning and meta-learning
- Overall, this paper presents a promising approach for test-time adaptation that can improve model performance without requiring re-training on each target domain.

## Summary

The paper "Back to the Source: Diffusion-Driven Test-Time Adaptation" proposes a novel approach for test-time adaptation that improves the accuracy of a model trained on source data when tested on shifted target data. Existing methods update the source model by re-training on each target domain, which is sensitive to the amount and order of the data and hyperparameters for optimization. Instead, this paper updates the target data by projecting all test inputs toward the source domain with a generative diffusion model. The proposed diffusion-driven adaptation method (DDA) shares its models for classification and generation across all domains, which are trained on the source domain and fixed during testing. The authors augment diffusion with image guidance and self-ensembling to automatically decide how much to adapt. The input adaptation by DDA is more robust than prior model adaptation approaches across a variety of corruptions, architectures, and data regimes on the ImageNet-C benchmark. With its input-wise updates, DDA succeeds where model adaptation degrades on too little data in small batches, dependent data in non-uniform order, or mixed data with multiple corruptions. The authors provide an extensive analysis of DDA's performance compared to existing state-of-the-art methods such as fine-tuning and meta-learning. Overall, this paper presents a promising approach for test-time adaptation that can improve model performance without requiring re-training on each target domain. It has practical implications for real world applications where models need to adapt quickly to new environments or datasets without access to large amounts of labeled training data.