Discriminator-Guided Adaptive Diffusion for Source-Free Test-Time Adaptation under Image Corruptions

Francesco Olivato, Cigdem Beyan, Vittorio Murino

Abstract. In this work, we study Source-Free Unsupervised Domain Adaptation under corruption-induced domain shifts, where performance degradation is caused by natural image corruptions that go beyond additive noise, including blur, weather effects, and digital artifacts. We propose a diffusion-based, input-level adaptation framework that operates entirely at test time and keeps all source-trained models frozen, explicitly targeting robustness to corrupted target inputs. Our method leverages a source-trained diffusion model as a generative prior and introduces a discriminator-guided adaptive diffusion strategy that dynamically controls the amount of perturbation applied to each test sample. Rather than relying on a fixed diffusion depth, the discriminator determines, on a per-image basis, when sufficient forward diffusion has been applied to suppress corruption-specific artifacts, with each corruption type effectively defining a distinct target domain. This adaptive stopping mechanism applies only the necessary amount of noise to remove domain-specific corruption while preserving class-discriminative structure. The reverse diffusion process then reconstructs a source-aligned image, optionally stabilized through structural guidance, which is classified using a frozen source-trained classifier. We evaluate the proposed approach across a broad spectrum of corruption-induced target domains, covering 15 diverse corruption types, and demonstrate more balanced robustness with competitive or improved performance across non-noise corruptions. Additional analyses reveal how the adaptive diffusion schedule responds to different corruption characteristics, highlighting the practicality, generality, and robustness of the proposed framework.

Install Dependencies

Create the environment and install dependencies via mamba:

mamba env create -f environment.yml
mamba activate dgadiffusion

ℹ️ Note: We used a Single NVIDIA GTX 4090 for the experiments.

Pretrained model

You can download the pretrained models used in our experiments here:

• our trained discriminator checkpoints
• the unconditional 256x256 diffusion model checkpoint used in our experiments

You can train your own discriminator by using the main training script (see Training section).

Get the Datasets

Download the original source dataset (e.g., ImageNet validation set) and the target corrupted dataset (e.g., ImageNet-C).

Project Structure

dgadiffusion
├─ checkpoints/
├─ guided_diffusion/
├─ images/
├─ dataset.py
├─ dataset_reconstruction.py
├─ environment.yml
├─ eval_reconstruction.py
├─ main.py
├─ model.py
├─ README.md
└─ utils.py

You have to clone the repo guided-diffusion to your local machine (as seen in project structure).

The checkpoints you downloaded must be placed under the checkpoints folder.

Experiment Tracking

This project uses Weights & Biases (WandB) for experiment tracking. Make sure to login to WandB before running the training script.

Training

The discriminator is trained to distinguish between noisy images originating from the target domain and noisy images drawn from the source distribution, effectively solving a binary classification task to identify domain-specific artifacts under varying levels of diffusion noise.

To train the discriminator, you can run the main training script. (Update with specific training arguments if needed):

python main.py \
    --wandb_entity <wandb_entity> \
    --dataset_source_dir <path_to_source_dataset> \
    --dataset_target_dir <path_to_target_dataset> \
    --dataset_target_domains <target_domain_1,target_domain_2,...> \
    --lr 0.00002 \
    --batch_size 64 \
    --epochs 5

Testing and Evaluation

At test time, the forward diffusion process is applied to generate a sequence of noisy representations. The pretrained discriminator monitors residual domain-specific cues and adaptively halts noising at the first timestep where its confidence drops below a predefined threshold. The subsequent reverse diffusion reconstructs a source-aligned image while minimizing semantic distortion.

To evaluate the adaptation on a specific target domain using this discriminator-guided adaptive stopping mechanism, use the eval_reconstruction.py script.

Based on the paper, the optimal configuration uses the custom evaluation type and a discriminator threshold of 0.5.

Run the following command, replacing the <...> placeholders with your specific paths and target domains:

python eval_reconstruction.py \
    --batch_size 8 \
    --dataset_target_domains <target_domain> \
    --eval_type custom \
    --checkpoint_path <path_to_checkpoint> \
    --dataset_source_dir <path_to_source_dataset> \
    --dataset_target_dir <path_to_target_dataset> \
    --disc_threshold 0.5

<target_domain> options for ImageNet-C: gaussian_noise, shot_noise, impulse_noise, defocus_blur, glass_blur, motion_blur, zoom_blur, snow, frost, fog, brightness, contrast, elastic_transform, pixelate, jpeg_compression.

Some Qualitative Examples

Citation

Please cite this work as follows if you find it useful:

@inproceedings{olivato2026discriminator,
  title={Discriminator-Guided Adaptive Diffusion for Source-Free Test-Time Adaptation under Image Corruptions},
  author={Olivato, Francesco and Beyan, Cigdem and Murino, Vittorio},
  booktitle={Proceedings of the 27th International Conference on Pattern Recognition (ICPR 2026)},
  year={2026}
}