pith. sign in

arxiv: 2604.23636 · v1 · submitted 2026-04-26 · 💻 cs.CV

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

Francesco Olivato , Cigdem Beyan , Vittorio Murino This is my paper

Pith reviewed 2026-05-08 06:45 UTC · model grok-4.3

classification 💻 cs.CV
keywords source-free domain adaptationtest-time adaptationdiffusion modelsimage corruptionsadaptive diffusiondiscriminator guidancerobustness
0
0 comments X

The pith

Discriminator-guided adaptive diffusion reconstructs each corrupted test image to match a frozen source model without retraining or source data access.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a test-time method to handle natural image corruptions such as blur, weather, and digital artifacts that degrade classifier performance. It starts with a source-trained diffusion model as a generative prior and adds noise to the input until a discriminator signals that corruption artifacts are suppressed, then reverses the diffusion to produce a clean reconstruction. The diffusion depth is chosen per image rather than fixed in advance, so that enough noise removes the specific corruption while class-discriminative structure remains intact. The resulting source-aligned image is then classified by the unchanged source model. This matters because it allows existing models to cope with varied real-world degradations at inference time without any additional training or labeled target data.

Core claim

The central claim is that a discriminator can determine, on a per-image basis, the precise forward diffusion depth needed to suppress corruption-specific artifacts while preserving class-discriminative structure; the subsequent reverse diffusion then produces a source-aligned image that a frozen source classifier can label accurately across multiple corruption-induced target domains.

What carries the argument

Discriminator-guided adaptive diffusion strategy, which uses discriminator output to select the forward diffusion steps applied to each test sample before reversing the process to reconstruct a source-aligned image.

If this is right

  • All source-trained models remain completely frozen, so no gradient updates or source data are required at test time.
  • The adaptive schedule produces more balanced accuracy across noise and non-noise corruptions than fixed-depth diffusion.
  • Each of the 15 corruption types elicits a distinct diffusion depth distribution, confirming that the mechanism responds to corruption-specific statistics.
  • Optional structural guidance during reverse diffusion can be added to further stabilize the reconstructed images without changing the core procedure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same per-image stopping logic could be tested on other forms of domain shift such as stylistic or sensor differences.
  • Efficiency gains would follow if the discriminator could be distilled into a lighter network for real-time use.
  • The observed correlation between corruption type and chosen diffusion depth supplies an empirical map that future work could use to predict required depth without running the discriminator.

Load-bearing premise

A source-trained diffusion model together with its paired discriminator can reliably detect when enough noise has been added to remove the corruption without erasing the semantic features needed for correct classification.

What would settle it

If replacing the adaptive per-image stopping rule with any fixed diffusion depth produces equal or higher classification accuracy on the same set of 15 corruption benchmarks, the advantage of the discriminator-guided schedule would be refuted.

Figures

Figures reproduced from arXiv: 2604.23636 by Cigdem Beyan, Francesco Olivato, Vittorio Murino.

Figure 1
Figure 1. Figure 1: Overview of the proposed discriminator-guided adaptive diffusion frame view at source ↗
Figure 2
Figure 2. Figure 2: (a) Illustration of discriminator-guided adaptive forward scheduling. For view at source ↗
Figure 3
Figure 3. Figure 3: Discriminator performance as a function of forward diffusion depth for view at source ↗
Figure 4
Figure 4. Figure 4: Representative examples of discriminator-guided adaptive stopping be view at source ↗
Figure 5
Figure 5. Figure 5: Discriminator-guided adaptive stopping behavior across all 15 corruption view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE visualization of feature representations for different corruption fam view at source ↗
read the original 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 domainspecific 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. The code is publicly available at https://github.com/fmolivato/dgadiffusion/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a source-free test-time adaptation framework (DGADiffusion) for robustness to image corruptions beyond additive noise. It keeps all source-trained models frozen and uses a source-trained diffusion model as a generative prior together with a discriminator that, on a per-image basis, selects the forward-diffusion depth needed to suppress corruption-specific artifacts while retaining class-discriminative structure; the reverse process then produces a source-aligned image that is classified by the frozen source classifier. The approach is evaluated across 15 corruption types and is claimed to yield more balanced robustness, with competitive or improved performance on non-noise corruptions (blur, weather, digital).

Significance. If the adaptive discriminator mechanism proves reliable, the work would supply a practical, input-level, test-time-only adaptation strategy that does not require source data access or target labels and that generalizes across heterogeneous corruption domains. The public code release supports reproducibility and would strengthen the contribution for the computer-vision community.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (method description): the central claim that the source-trained discriminator can, without target labels or clean references, reliably identify the precise forward-diffusion stopping point that removes corruption artifacts while preserving semantics is load-bearing. The provided text offers no quantitative evidence (e.g., discriminator accuracy per corruption, ablation on depth-selection error, or failure-case analysis) that this mapping generalizes from source data to non-noise corruptions; the skeptic concern therefore remains unaddressed and directly affects the validity of the adaptive schedule.
  2. [§5] §5 (experiments): the claim of “competitive or improved performance across non-noise corruptions” is stated without the concrete metrics, baseline comparisons, or per-corruption tables that would allow assessment of effect size or statistical significance. Because the abstract already notes the absence of these details, the evidential support for the headline result is currently insufficient.
minor comments (2)
  1. [Abstract] The abstract would benefit from a single sentence summarizing the quantitative gains (e.g., average accuracy improvement on the 15 corruptions) to give readers an immediate sense of the result magnitude.
  2. [§4] Notation for the adaptive stopping criterion (e.g., the discriminator output threshold or the mapping from image statistics to diffusion timestep) should be introduced with an equation in the method section for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger quantitative support. We address each major comment below and will revise the manuscript to incorporate the requested evidence and details.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (method description): the central claim that the source-trained discriminator can, without target labels or clean references, reliably identify the precise forward-diffusion stopping point that removes corruption artifacts while preserving semantics is load-bearing. The provided text offers no quantitative evidence (e.g., discriminator accuracy per corruption, ablation on depth-selection error, or failure-case analysis) that this mapping generalizes from source data to non-noise corruptions; the skeptic concern therefore remains unaddressed and directly affects the validity of the adaptive schedule.

    Authors: We agree that direct quantitative validation of the discriminator's depth-selection behavior is essential for the central claim. The current manuscript describes the adaptive mechanism in §4 and notes additional analyses of the diffusion schedule's response to corruption characteristics, but does not report explicit per-corruption discriminator accuracy, depth-selection error ablations, or dedicated failure-case studies. In the revision we will add these elements, including quantitative metrics on how well the discriminator generalizes to non-noise corruptions and an analysis of selection errors. revision: yes

  2. Referee: [§5] §5 (experiments): the claim of “competitive or improved performance across non-noise corruptions” is stated without the concrete metrics, baseline comparisons, or per-corruption tables that would allow assessment of effect size or statistical significance. Because the abstract already notes the absence of these details, the evidential support for the headline result is currently insufficient.

    Authors: We acknowledge that the experimental claims would be more convincing with granular data. While the manuscript evaluates the method across 15 corruption types and reports overall balanced robustness, it does not present full per-corruption tables with baseline comparisons or effect-size statistics in the main text. In the revision we will expand §5 with detailed per-corruption performance tables, including comparisons to relevant baselines, to enable direct assessment of improvements on non-noise corruptions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is a novel proposal with external evaluation

full rationale

The paper introduces a new source-free test-time adaptation framework using a source-trained diffusion model as generative prior and a discriminator to adaptively select per-image diffusion depth for corruption suppression. The derivation chain consists of proposing the adaptive stopping mechanism, applying reverse diffusion, and evaluating on 15 corruption types with a frozen classifier. No equations, parameters, or central claims reduce to self-definition, fitted inputs renamed as predictions, or load-bearing self-citations. The discriminator's role is defined externally via its training on source data and tested on target corruptions without tautological reduction. This is a standard architectural proposal whose validity rests on empirical results rather than internal equivalence to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard diffusion model assumptions and domain adaptation premises without introducing new free parameters or invented entities in the abstract description.

axioms (2)
  • domain assumption A source-trained diffusion model can act as a generative prior to reconstruct source-aligned images from corrupted inputs.
    Invoked as the basis for the reverse diffusion reconstruction step.
  • domain assumption A discriminator can accurately detect when corruption artifacts are suppressed without loss of class-discriminative information.
    Central to the adaptive stopping mechanism described.

pith-pipeline@v0.9.0 · 5567 in / 1254 out tokens · 50480 ms · 2026-05-08T06:45:32.243307+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

  1. [1]

    Source-free adaptation to measurement shift via bottom-up feature restoration

    Cian Eastwood, Ian Mason, Chris Williams, and Bernhard Schölkopf. Source-free adaptation to measurement shift via bottom-up feature restoration. InICLR, 2021

    2021

  2. [2]

    Source-free unsupervised domain adaptation: A survey.Neural Networks, 174:106230, 2024

    Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey.Neural Networks, 174:106230, 2024

    2024

  3. [3]

    Back to the source: Diffusion-driven adaptation to test-time corruption

    Jin Gao, Jialing Zhang, Xihui Liu, et al. Back to the source: Diffusion-driven adaptation to test-time corruption. InCVPR, pages 11786–11796, 2023

    2023

  4. [4]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. InICCV, 2021

    2021

  5. [5]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. InICLR, 2019. Adaptive Diffusion for Source-Free Test-Time Adaptation 17 (a) Noise corruptions (b) Blur corruptions (c) Weather corruptions (d) Digital corruptions Fig.6:t-SNEvisualizationoffeaturerepresentationsfordifferentcorruptionfam- i...

    2019

  6. [6]

    Denoising diffusion probabilistic mod- els

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic mod- els. InNeurIPS, 2020

    2020

  7. [7]

    Cycada: Cycle-consistent adver- sarial domain adaptation

    Judy Hoffman, Eric Tzeng, Taesung Park, et al. Cycada: Cycle-consistent adver- sarial domain adaptation. InICML, 2018

    2018

  8. [8]

    Context-aware pseudo- label refinement for source-free domain adaptive fundus image segmentation

    Zheang Huai, Xinpeng Ding, Yi Li, and Xiaomeng Li. Context-aware pseudo- label refinement for source-free domain adaptive fundus image segmentation. In MICCAI, pages 618–628, 2023

    2023

  9. [9]

    Test-time classifier adjustment module for model-agnostic domain generalization

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. InNeurIPS, 2021

    2021

  10. [10]

    Venkatesh Babu, et al

    Jogendra Nath Kundu, Naveen Venkat, R. Venkatesh Babu, et al. Universal source- free domain adaptation. InCVPR, 2020

    2020

  11. [11]

    Modeladaptation: Unsupervised domain adaptation without source data

    RuiLi,QianfenJiao,WenmingCao,Hau-SanWong,andSiWu. Modeladaptation: Unsupervised domain adaptation without source data. InCVPR, 2020. 18 F. Olivato et al

    2020

  12. [12]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. InICML, pages 6028–6039, 2020

    2020

  13. [13]

    Swin transformer: Hierarchical vision trans- former using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, et al. Swin transformer: Hierarchical vision trans- former using shifted windows. InICCV, 2021

    2021

  14. [14]

    Diffusion models for adversarial purification

    Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. InICML, 2022

    2022

  15. [15]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, et al. Efficient test-time model adaptation without forgetting. InICML, 2022

    2022

  16. [16]

    Towards stable test-time adaptation in dynamic wild world.arXiv preprint arXiv:2302.12400, 2023

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, et al. Towards stable test-time adap- tation in dynamic wild world. InarXiv preprint arXiv:2302.12400, 2023

    work page arXiv 2023

  17. [17]

    Generalization on unseen domains via inference-time label-preserving target projections

    Prashant Pandey, Mrigank Raman, et al. Generalization on unseen domains via inference-time label-preserving target projections. InCVPR, 2021

    2021

  18. [18]

    Uncertainty-induced transferabil- ity representation for source-free unsupervised domain adaptation.IEEE TIP, 32:2033–2048, 2023

    Jiangbo Pei, Zhuqing Jiang, Aidong Men, et al. Uncertainty-induced transferabil- ity representation for source-free unsupervised domain adaptation.IEEE TIP, 32:2033–2048, 2023

    2033

  19. [19]

    Unsupervised domain adaptation via domain-adaptive diffusion.IEEE TIP, 2024

    Duo Peng, Qiuhong Ke, ArulMurugan Ambikapathi, et al. Unsupervised domain adaptation via domain-adaptive diffusion.IEEE TIP, 2024

    2024

  20. [20]

    Zimmermann, et al

    Evgenia Rusak, Lukas Schott, Roland S. Zimmermann, et al. A simple way to make neural networks robust against diverse image corruptions. InECCV, 2020

    2020

  21. [21]

    Image super-resolution via iterative refinement.IEEE TPAMI, 45:4713–4726, 2022

    Chitwan Saharia, Jonathan Ho, William Chan, et al. Image super-resolution via iterative refinement.IEEE TPAMI, 45:4713–4726, 2022

    2022

  22. [22]

    Improving robustness against common corruptions by covariate shift adaptation

    Steffen Schneider, Evgenia Rusak, Luisa Eck, et al. Improving robustness against common corruptions by covariate shift adaptation. InNeurIPS, 2020

    2020

  23. [23]

    Sudre, et al

    Thomas Varsavsky, Mauricio Orbes-Arteaga, Carole H. Sudre, et al. Test-time unsupervised domain adaptation. InMICCAI, 2020

    2020

  24. [24]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, et al. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021

    2021

  25. [25]

    Memo: Test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. InNeurIPS, 2021

    2021