Research map / Label-efficient and robust learning / Test-time adaptation

GDA: Generalized Diffusion for Robust Test-time Adaptation

CVPR 2024 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GDA, a diffusion-based test-time adaptation method that guides DDPM/DDIM reverse sampling with marginal entropy, CLIP style, and patch-wise contrastive content losses to shift OOD images back to the source domain without changing classifier weights.

Problem

Test-time adaptation for out-of-distribution images under diverse shifts (corruptions, renditions, sketches, stylized images); prior diffusion-based adaptation mainly handled pixel-level corruptions and high-frequency noise.

Approach

Structural guidance during diffusion reverse sampling that combines a classifier-derived marginal entropy loss (over augmented samples) with CLIP-based style loss and contrastive content-preservation loss, plus confidence-based filtering between original and adapted samples.

Key results

• On ImageNet-C severity 3, GDA reaches 41.7% (ResNet50), 65.24% (ConvNext-T), 59.35% (Swin-T) vs. standard 37.30, 59.60, 54.33.
• On Rendition with ResNet50, GDA achieves 44.5% vs. DDA 42.0% and standard 37.0%; on Sketch with Swin-T GDA reaches 35.9% vs. standard 29.0%.
• GDA runs about 3.49 s per sample on Rendition/ResNet50 versus 31.7 s for Diffpure and 13.5 s for DDA50, with 16 augmentations and 50 time steps.

Measured effects reported in the paper

Surprising findings

• Diffpure substantially degrades accuracy on ImageNet-C (e.g., 15.83% on ResNet50 vs. 37.30% standard).
• DDA and Diffpure do not improve, and sometimes hurt, accuracy on the Sketch benchmark.

Limitations

• Adaptation is per-sample and requires running a diffusion reverse process, adding substantial inference cost compared with standard inference.
• Gains on some benchmarks (e.g., Stylized for ResNet50 comparisons) are modest and the method was evaluated only on ImageNet-derived OOD benchmarks and image classification.
• Accuracy saturates with more augmentations and the method depends on hyperparameter (Lagrange multiplier) choices per benchmark.

Experimental setup

Datasets: ImageNet ImageNet-C ImageNet-Rendition ImageNet-Sketch ImageNet-Stylized

Benchmarks: ImageNet-C (severity 3) ImageNet-Rendition ImageNet-Sketch ImageNet-Stylized

Metrics: classification accuracy run time per sample entropy loss distribution

Baselines: Standard (no adaptation) DDA Diffpure w/o marginal

Builds on: DDA DDPM DDIM Diffpure MEMO CUT contrastive unpaired translation

Closest papers by meaning

• GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.86)
• Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.85)
• GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models (ICML 2025 · similarity 0.85)
• In Search of Forgotten Domain Generalization (ICLR 2025 · similarity 0.85)
• Test-Time Degradation Adaptation for Open-Set Image Restoration (ICML 2024 · similarity 0.85)
• Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.85)