Paper deep dive
Distilling Image Prototypes for Guided Test-Time Adaptation
Liwen Wang, Xingbo Dong, Iman Yi Liao, Deyin Liu, Massimo Tistarelli, Lin Yuanbo Wu, Zhe Jin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/10/2026, 3:16:07 AM
Summary
The paper introduces DIPTTA, a framework for Test-Time Adaptation (TTA) that addresses error accumulation and catastrophic forgetting. It utilizes a Distilling Image Prototype (DIP), a compact set of synthetic images, to dynamically regenerate feature prototypes aligned with the evolving model state, preventing forgetting. Additionally, it employs Source-Calibrated Uncertainty Estimation (SCUE) using the DIP to debias uncertainty quantification and suppress error accumulation from noisy pseudo-labels.
Entities (7)
Relation Signals (6)
DIPTTA → uses → Distilling Image Prototype
confidence 98% · The core of the proposed approach is the introduction of a Distill Image Prototype (DIP)...
Distilling Image Prototype → enables → Dynamic Feature Prototype Replay
confidence 95% · This prototype enables a dynamic feature replay mechanism that continuously generates feature prototypes aligned with the current state of the model...
Distilling Image Prototype → enables → Source-Calibrated Uncertainty Estimation
confidence 95% · Furthermore, the DIP anchors a source-calibrated uncertainty estimation method...
Dynamic Feature Prototype Replay → prevents → Catastrophic Forgetting
confidence 93% · ...thus effectively preventing catastrophic forgetting.
Source-Calibrated Uncertainty Estimation → suppresses → Error Accumulation
confidence 93% · ...thereby robustly suppressing error accumulation.
DIPTTA → solves → Test-Time Adaptation
confidence 90% · To address these issues, this paper proposes a novel framework, Distilling Image Prototype for Guided Test-Time Adaptation (DIPTTA).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Test-Time Adaptation (TTA) enhances the robustness of models against distribution shifts but faces two critical challenges: error accumulation from noisy pseudo-labels and catastrophic forgetting of source knowledge. Uncertainty-based approaches designed to mitigate error accumulation often yield overconfident or computationally expensive estimates, while strategies intended to prevent forgetting via prototype replay rely on static representations that easily become misaligned as the model adapts. To address these issues, this paper proposes a novel framework, Distilling Image Prototype for Guided Test-Time Adaptation (DIPTTA). The core of the proposed approach is the introduction of a Distill Image Prototype (DIP), a compact set of synthetic images that serves as a dynamic and regenerative anchor of source knowledge. This prototype enables a dynamic feature replay mechanism that continuously generates feature prototypes aligned with the current state of the model, thus effectively preventing catastrophic forgetting. Furthermore, the DIP anchors a source-calibrated uncertainty estimation method, which provides a less biased measure of sample reliability by leveraging stable source knowledge, thereby robustly suppressing error accumulation. Extensive experiments on multiple benchmarks demonstrate that DIPTTA significantly outperforms state-of-the-art methods, particularly under severe domain shifts. The source code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.09737v1
- Canonical: https://arxiv.org/abs/2609.09737v1
Trouble viewing inline? Open PDF directly →
Full Text
67,929 characters extracted from source content.
Expand or collapse full text
Distilling Image Prototypes for Guided Test-Time AdaptationThanks: This work was supported by the National Natural Science Foundation of China (Grant No. 62306003), the Open Research Fund of Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) (Grant No. GML-KF-24-29), and the Open Foundation of Jiangxi Provincial Key Laboratory of Image Processing and Pattern Recognition (Grant No. ET202404437). (Corresponding author: Xingbo Dong and Zhe Jin)Thanks: Liwen Wang, Xingbo Dong, Deyin Liu, and Zhe Jin are with the Anhui Provincial Key Laboratory of Secure Artificial Intelligence, Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Artificial Intelligence, Anhui University, Hefei 230093, China (e-mail: liwenwang@stu.ahu.edu.cn; xingbo.dong@ahu.edu.cn; jinzhe@ahu.edu.cn). Iman Yi Liao is with the School of Computer Science, University of Nottingham (Malaysia Campus), Selangor Darul Ehsan, 43500, Malaysia (e-mail: Iman.Liao@nottingham.edu.my). Massimo Tistarelli is with the Department of Engineering, University of Sassari, Sassari, Italy (e-mail: tista@uniss.it). Lin Yuanbo Wu is with the School of Engineering, University of Warwick, CV4 7AL Coventry, U.K. (e-mail: jolin.lwu@gmail.com). PubID: pubid: 0000–0000/00$00.00 © 2026 IEEE Liwen Wang Xingbo Dong Iman Yi Liao Deyin Liu Affiliation: Massimo Tistarelli, Lin Yuanbo Wu, , Zhe Jin, Abstract Test-Time Adaptation (TTA) enhances the robustness of models against distribution shifts but faces two critical challenges: error accumulation from noisy pseudo-labels and catastrophic forgetting of source knowledge. Uncertainty-based approaches designed to mitigate error accumulation often yield overconfident or computationally expensive estimates, while strategies intended to prevent forgetting via prototype replay rely on static representations that easily become misaligned as the model adapts. To address these issues, this paper proposes a novel framework, Distilling Image Prototype for Guided Test-Time Adaptation (DIPTTA). The core of the proposed approach is the introduction of a Distill Image Prototype (DIP), a compact set of synthetic images that serves as a dynamic and regenerative anchor of source knowledge. This prototype enables a dynamic feature replay mechanism that continuously generates feature prototypes aligned with the current state of the model, thus effectively preventing catastrophic forgetting. Furthermore, the DIP anchors a source-calibrated uncertainty estimation method, which provides a less biased measure of sample reliability by leveraging stable source knowledge, thereby robustly suppressing error accumulation. Extensive experiments on multiple benchmarks demonstrate that DIPTTA significantly outperforms state-of-the-art methods, particularly under severe domain shifts. The source code is available at https://github.com/LiwenWang919/DIPTTA. Index Terms: Test-time adaptation, distilling image prototype, dynamic replay, source-calibrated uncertainty estimation. I Introduction Deep learning models struggle when test data shifts from their training distribution. Test-Time Adaptation (TTA) addresses this by enabling online adaptation using unlabeled data [1, 2, 3, 4]. However, TTA faces two major hurdles: error accumulation from noisy pseudo-labels and the catastrophic forgetting of original source knowledge. Current strategies to combat catastrophic forgetting often involve replaying information from the source domain. A prominent line of work [5, 6] advocates for replaying pre-computed feature prototypes, which are class-wise average feature vectors derived from the source data. This line of research also includes methods that maintain a source-like feature bank or use nearest-neighbor retrieval [7]. However, these approaches have a critical flaw: these prototypes are computed once using the pretrained source model and remain static throughout the adaptation process. As the model updates its parameters to the target domain, its internal feature representation evolves, causing a growing misalignment between the static prototypes and the current feature space of the model. Consequently, these outdated prototypes provide a progressively potentially misleading anchor, thereby failing to prevent forgetting effectively. This fundamental limitation motivates the first research question: How can a replay-capable knowledge anchor be established that dynamically evolves with the model to alleviate catastrophic forgetting in a continually shifting environment? Conversely, mitigating error accumulation typically relies on uncertainty estimation to identify and down-weight unreliable predictions. However, existing methods face a significant dilemma. Entropy-based techniques [8, 9, 10] are prone to overconfidence, especially on out-of-distribution data, as they only capture the sharpness of the output distribution, not the true epistemic uncertainty of the model [1, 11, 2, 12]. While more principled Bayesian methods [13, 14] can capture this uncertainty, they are frequently computationally prohibitive for online TTA. Recent efforts to improve their efficiency, such as through last-layer variational inference [15] or simplified Laplace approximations [16], still face challenges in non-stationary TTA settings. More critically, both types of estimates are inherently biased in the TTA context because they are computed solely based on the reaction of the model to the target data, which is itself non-stationary and potentially noisy. There is no stable reference point from the source domain to ground these estimates. This leads to the second research question: How can a stable source-domain anchor be leveraged to achieve unbiased uncertainty estimation, thereby reliably suppressing error accumulation? It is identified that the root cause of both limitations is the lack of a dynamic, reliable, and efficient source of knowledge during adaptation. To bridge this gap, this paper introduces a concept dubbed Distilling Image Prototype (DIP). A DIP is not a static feature vector but a compact set of synthetic images distilled from the entire source dataset. This distillation process refines synthetic images so that a model trained on them closely resembles the performance of a model trained on the complete source data [17]. The key advantage of the DIP is its regenerability: it exists in the image space, allowing it to be continuously processed through the updated feature extractor during adaptation. This enables the on-demand generation of feature prototypes that are inherently aligned with the current feature space of the model, thereby solving the misalignment issue of static prototypes. Guided by this concept, the Distilling Image Prototype for Guided Test-Time Adaptation (DIPTTA) framework is proposed. The DIP acts as a unified anchor, enabling two key innovations. First, to establish a replay-capable knowledge anchor that dynamically evolves with the model to prevent catastrophic forgetting in a continually shifting environment, the feature prototypes are regenerated for each incoming batch of target data by forwarding the DIP through the current student model. These dynamic prototypes are then used in a contrastive loss to pull the features of pseudo-labeled target samples towards their corresponding class anchors, ensuring that the model retains source knowledge without feature space distortion. Second, to leverage a stable source-domain anchor to achieve unbiased uncertainty estimation, thereby reliably suppressing error accumulation, the predictive uncertainty is computed by observing the response of the model (specifically, the classifier head) to the feature prototypes generated from the DIP. This measures the consistency between the target sample and the stable source knowledge, rather than the self-confidence of the model regarding the target sample alone. The resulting uncertainty measure is inherently unbiased and serves as a robust weight to modulate the contribution of each sample to the adaptation loss, effectively curating the learning signal. The key contributions can be summarized as follows: • We introduce the novel concept of a Distilling Image Prototype (DIP) for test-time adaptation. A DIP is a compact set of synthetic images that serve as a dynamic and regenerative knowledge anchor, overcoming the rigidity of static source representations and providing a principled solution to knowledge preservation in non-stationary environments. • We develop a Dynamic Feature Prototype Replay (DFPR) mechanism based on the DIP. This mechanism continuously regenerates feature prototypes aligned with the model’s current state, effectively addressing catastrophic forgetting in continual test-time adaptation without access to the data of the source domain. • We propose a Source-Calibrated Uncertainty Estimation (SCUE) method that leverages the DIP to debias uncertainty quantification. By anchoring the uncertainty measure in the stable source knowledge, our method provides a more reliable signal for weighting target samples, leading to robust adaptation against error accumulation. • Extensive experiments on multiple benchmarks demonstrate that our proposed DIP framework achieves state-of-the-art performance. The consistent improvements across various scenarios, especially under strong domain shifts, validate the effectiveness and robustness of our approach. I Related Work I-A Test Time Adaptation Test-Time Adaptation (TTA) methods aim to adapt models via objectives applied directly to the test data. A dominant strategy has been output-level adaptation, which includes entropy minimization [8, 9, 18] to encourage confident predictions, and self-training [19, 20, 21, 7], which uses high-confidence pseudo-labels for supervision. These methods are highly susceptible to error accumulation from noisy pseudo-labels, especially under significant domain shifts. To enhance robustness, another line of work focuses on feature-level adaptation [22, 23, 24, 25]. This includes methods based on consistency regularization, such as the Mean Teacher paradigm [26, 24, 27, 28], which stabilizes predictions by enforcing agreement between a teacher model and a student model under different augmentations. Furthermore, contrastive learning techniques [29, 30] have been adopted to learn a structured feature space with compact class clusters, improving feature discrimination in the target domain. While these methods promote feature invariance, they still fundamentally rely on the quality of pseudo-labels or the predictions of the teacher model, which can be unreliable for OOD samples [5]. A critical challenge in TTA, particularly in the continual setting (CTTA) [31, 32], is catastrophic forgetting. To address this without accessing source data, recent methods have explored replaying source knowledge [33]. Notably, several approaches [5, 6] replay pre-computed, static feature prototypes of the source classes. However, a fundamental limitation persists: as the model adapts, its feature representation evolves, causing these statically computed prototypes to become progressively misaligned with the current feature space [34]. This misalignment undermines their effectiveness as a stable knowledge anchor. Dataset distillation is rapidly becoming a research hotspot in machine learning, motivated primarily by the increasing scale of data and substantial model training overhead. By generating small yet potent synthetic datasets, this technology paves the way for more efficient learning and has demonstrated considerable value in numerous scenarios [35], including its application in continual learning [36, 37]. In this paper, dataset distillation is introduced into TTA to alleviate the problem of catastrophic forgetting; a replay mechanism for latent prototypes driven by dataset distillation is established, and its feasibility is verified through experiments. The proposed DIPTTA framework addresses the above limitations through a novel paradigm centered on a Distilling Image Prototype (DIP). Unlike static feature prototypes, the DIP is a set of synthetic images that allows for the dynamic regeneration of feature prototypes aligned with the evolving model, effectively overcoming the misalignment issue. This dynamic replay mechanism is coupled with a source-calibrated uncertainty estimation method that leverages the DIP to debias uncertainty quantification, offering a more robust solution to both catastrophic forgetting and error accumulation. I-B Uncertainty Quantification in Test Time Adaptation Uncertainty estimation is crucial in TTA to identify unreliable predictions under domain shifts and to prevent error accumulation [38, 39, 40]. Existing methods can be broadly categorized by their underlying principles. Simple predictive metrics, such as the prediction entropy utilized in methods like Tent [8] and SHOT [41], are computationally lightweight. However, these approaches primarily reflect the sharpness of the output distribution and often fail to account for the epistemic uncertainty of the model. This leads to overconfident and miscalibrated predictions on out-of-distribution data. Principled Bayesian methods offer a more formal framework for capturing epistemic uncertainty. Techniques such as MC Dropout [13], model ensembles [42, 43], and Bayesian neural networks [44, 14] model distributions over parameters or predictions. While these methods can provide better-calibrated uncertainty estimates, their computational demands often hinder their application in online TTA scenarios. Other lightweight strategies include using the maximum softmax probability [45] or energy scores [46] as uncertainty proxies. A more fundamental limitation common to both categories is that they estimate uncertainty based solely on the target domain and the adapting model. The model itself is changing, and the target data is non-stationary. This lack of a stable reference point means that the uncertainty estimates can be inherently biased by the shifting target distribution, lacking calibration to the source knowledge [47, 48]. This limitation is addressed by introducing a source-calibrated uncertainty framework. Instead of relying exclusively on the output of the adapting model for a target sample, the Distilling Image Prototype (DIP) is leveraged as a stable anchor representing source knowledge. This enables the estimation of the consistency of a target sample with the source domain, resulting in a less biased measure of reliability. By grounding the uncertainty quantification in the stable source representation, the proposed method provides a more robust mechanism for identifying unreliable predictions during adaptation. Fig. 1: Overall framework of our proposed DIPTTA. The top part visualizes the Distilling Image Prototype (DIP) module, which generates high-fidelity synthetic prototypes via a two-loop distillation strategy. The bottom part depicts the core online adaptation pipeline, where a student model is continuously updated with the guidance of a teacher model. Specifically, a source-calibrated uncertainty estimation and reweighting mechanism is embedded to dynamically calibrate the gradient contribution from target data, effectively mitigating distribution shift issues in open-world adaptation scenarios. I Method In this section, we detail our method, Distill Image Prototype-guided Test-Time Adaptation (DIPTTA). I-A Notation and Model Architecture Let =(xi,yi)D_S=\(x_i^S,y_i^S)\ be the source dataset, with samples from a source domain distribution P(x,y)P_S(x,y). Similarly, =(xi)D_T=\(x_i^T)\ is the unlabeled target dataset from a target domain distribution P(x)P_T(x). The fundamental challenge in this setting is that the source distribution is different from the target distribution, i.e., P(x,y)≠P(x)P(y|x)P_S(x,y)≠ P_T(x)P(y|x), although they share a common label space Y. Given a well-trained model M_S pretrained on the full source data D_S, the goal is to adapt M_S to the target distribution under the test-time adaptation setting, which adapts the model using sequential unlabeled target batches without access to source data. I-A1 Model Architecture To achieve this, a mean teacher framework is employed. At the start (k=0k=0), both the student model (Mst=hst(fst(⋅))M_st=h_st(f_st(·))) and the teacher model (Mte=hte(fte(⋅))M_te=h_te(f_te(·))) are initialized with the parameters pretrained in the source domain, such that Mst,k=0=Mte,k=0=M_st,k=0=M_te,k=0=M_S. For each incoming batch BkB_k, the parameters of the student model θst _st are optimized. Subsequently, the parameters of the teacher model θte _te are updated via an Exponential Moving Average (EMA) of the parameters of the student model: Mte←αMte+(1−α)MstM_te←α M_te+(1-α)M_st (1) This framework enables the model to dynamically adapt to non-stationary domains by utilizing a stream of unlabeled data to track distributional changes, while concurrently receiving stable supervisory signals from the teacher component. I-B Dynamic Feature Prototype Replay To overcome the inherent limitations of static feature prototypes, the knowledge anchor is shifted from the feature space to the image space. Unlike existing methods that rely on frozen feature vectors, the DIP framework optimizes a set of synthetic images VP_V to encapsulate the essential geometry of the source distribution. This design choice is grounded in the principle of representation invariance: while the feature extractor fstf_st undergoes significant drift during target adaptation, the semantic content within the image space remains a constant reference point. By regenerating features through the latest model state (Eq. 6), it is ensured that the source knowledge anchor is consistently aligned with the current model, thus theoretically eliminating the misalignment error that plagues conventional replay mechanisms. I-B1 Distilling Image Prototype Initialization and Optimization Before optimization, a small set of synthetic images =(s1,y1),…,(sn,yn)P_V=\(s_1,y_1),…,(s_n,y_n)\ with n≪||n |D_S| is randomly initialized and subsequently optimized for each class. This procedure is a bi-level optimization process. In the Simulated Training (Inner Loop), a virtual model M_V (a copy of M_S) is updated using a mini-batch from P_V to minimize a simulated training loss, typically on the classifier head or a few top layers, thereby mimicking lightweight adaptation. The inner loop optimization can be formulated as: θ(t+1)=θ(t)−α∇θℒinner(Bν,θ(t)),θ^(t+1)=θ^(t)-α _θL_inner(B_ν;θ^(t)), (2) where θ(t)θ^(t) represents the model parameters at inner loop step t, α is the learning rate of the inner loop, BνB_ν is a mini-batch sampled from the synthetic dataset P_V, and ℒinnerL_inner is typically the cross-entropy loss defined as ℒinner=1|Bν|∑(si,yi)∈BνCE(M(si),yi).L_inner= 1|B_ν| _(s_i,y_i)∈ B_νCE(M_V(s_i),y_i). (3) In the Meta-Optimization (Outer Loop), the synthetic images sis_i in P_V are optimized based on the performance of the virtually trained M_V on a mini-batch from the source dataset D_S. The meta-loss ℒmetaL_meta drives the synthetic images to encapsulate the essential knowledge of D_S. The outer loop optimization can be formulated as: si←si−β∇siℒmeta(B,θ∗(si)),s_i← s_i-β _s_iL_meta(B_S;θ^*(s_i)), (4) where sis_i denotes the synthetic images being optimized, β is the outer loop learning rate, θ∗(si)θ^*(s_i) represents the model parameters after inner loop training on P_V, B_S is a mini-batch sampled from the source dataset D_S, and ℒmetaL_meta is generally the cross-entropy loss defined as ℒmeta=1|B|∑(xj,yj)∈BCE(M(xj),yj).L_meta= 1|B_S| _(x_j,y_j)∈ B_SCE(M_V(x_j),y_j). (5) This optimization can include matching feature distributions between M_V on P_V and M_S on D_S, and ensuring classifier consistency. Optional privacy-friendly regularization (pixel value constraints and smoothness regularization) can be applied during optimization to avoid generating images that are highly similar to the original samples. I-B2 Dynamic Feature Prototype Generation After optimization, the final synthetic images ∗P_V^* are obtained. During test-time adaptation, feature prototypes are dynamically generated by passing ∗P_V^* through the feature extractor fst(k)f_st^(k) of the current student model at each adaptation step k. For each class c, the feature prototype f(k)[c]P_f^(k)[c] is computed as: f(k)[c]=Mean(fst(k)(s)∣(s,y)∈∗,y=c)P_f^(k)[c]=Mean(\f_st^(k)(s) (s,y) _V^*,y=c\) (6) These dynamically generated feature prototypes f(k)P_f^(k) are utilized in the proposed contrastive learning strategy. Given the features fst(xi)f_st(x_i) extracted by the student model for a target sample xix_i, the similarity between fst(xi)f_st(x_i) and all feature prototypes is computed. The contrastive loss for a single sample xix_i is defined as: ℒcontrastive(xi)= _contrastive(x_i)= −logexp(sim(fst(xi),f(k)[y^i])/τ)∑c∈Cexp(sim(fst(xi),f(k)[c])/τ), - (sim(f_st(x_i),P_f^(k)[ y_i])/τ) _c∈ C (sim(f_st(x_i),P_f^(k)[c])/τ), (7) where y^i y_i is the pseudo-label predicted by the teacher model for xix_i, sim(⋅,⋅)sim(·,·) denotes cosine similarity, and τ is a temperature parameter. This contrastive learning approach alleviates feature uncertainty in the target domain by utilizing dynamically updated feature prototypes f(k)P_f^(k). The prototypes are driven by the DIP ∗P_V^*, which serves as a compact substitute for source data. The contrastive loss guides the model by drawing the features of target samples closer to their assigned class prototypes while pushing them away from other prototypes, effectively anchoring the target feature space to robust source knowledge. I-C Source-Calibrated Uncertainty Estimation To mitigate error accumulation from noisy pseudo-labels, a source-calibrated uncertainty estimation method is proposed. Unlike traditional approaches that rely solely on target data, the proposed method uses the DIP as a stable anchor from the source domain to provide unbiased uncertainty quantification. The source-calibrated uncertainty stems from the modeling of uncertainty sources. Traditional methods estimate uncertainty based solely on p(y|xt,θ)p(y|x_t,θ), which conflates aleatoric uncertainty (from target data noise) and epistemic uncertainty (from model ignorance due to domain shift). During Test-Time Adaptation (TTA), parameters θ evolve rapidly, making epistemic uncertainty difficult to isolate. The proposed method introduces a source anchor zsz_s (from DIP) to disentangle this conflation by evaluating the conditional distribution p(y|xt,zs,θ)p(y|x_t,z_s,θ). Within a Bayesian framework: p(y|xt,zs,θ)∝p(y|xt,θ)⋅p(zs|y,xt,θ)p(y|x_t,z_s,θ) p(y|x_t,θ)· p(z_s|y,x_t,θ) (8) Here, p(zs|y,xt,θ)p(z_s|y,x_t,θ) serves as a calibration factor, measuring the plausibility of the source anchor zsz_s given the target sample and its prediction. If a target sample has high prediction confidence p(y|xt,θ)p(y|x_t,θ) but its features diverge from the source anchor (low p(zs|y,xt,θ)p(z_s|y,x_t,θ)), the calibrated probability p(y|xt,zs,θ)p(y|x_t,z_s,θ) decreases, indicating potential miscalibration. Conversely, good alignment with the source anchor increases the calibrated confidence [49]. Thus, the source-calibrated uncertainty, (xt)=ℍ[p(y|xt,zs,θ)]U(x_t)=H[p(y|x_t,z_s,θ)], corrects the biased distribution from target data alone, providing a more reliable measure of predictive uncertainty that is robust to domain shift. While Eq. 8 defines the ideal source-calibrated distribution, directly computing the integral over the entire parameter space is computationally intractable for online TTA. To bridge this gap, the response of the model to the DIP-based prototypes is treated as a proxy for epistemic uncertainty. Specifically, the focus is placed on the uncertainty of the classification head θh _h, as the feature extractor fstf_st is assumed to be a deterministic mapping. By applying a Laplace Approximation (LA) centered at the MAP estimate θMAP _MAP, the abstract calibration problem is transformed into a practical weight estimation task (Eq. 11), thereby anchoring the target prediction reliability in the stability of the pre-acquired source knowledge. A direct implementation of the aforementioned theory is difficult. Therefore, the paper adopts an approximation method based on Bayesian inference, which focuses on modeling the uncertainty of the model parameters and thereby indirectly achieves source domain calibration. The posterior distribution over the network parameters θ, conditioned on the source knowledge represented by the DIP-based prototypes fP_f, is given by p(θ|f)∝p(θ)p(f|θ)p(θ|P_f) p(θ)p(P_f|θ). As computing this posterior is intractable, a Laplace Approximation (LA) is employed, which forms a Gaussian distribution around a mode of the posterior, θMAP _MAP: p(θ|f)≈(θ|θMAP,H−1)p(θ|P_f) (θ| _MAP,H^-1) (9) Here, H is the negative Hessian of the log-posterior evaluated at θMAP _MAP. To maintain computational feasibility, this Bayesian treatment is applied only to the final classification layer of the student model, hsth_st, using the last-layer Laplace approximation. The feature extractor fstf_st remains deterministic. For efficiency, the Hessian is approximated using Kronecker-factored Laplace Approximation (KFLA) [50] where H≈V⊗UH≈ V U. This enables the efficient estimation of the predictive posterior distribution for a given feature representation z: p(y|z,f)≈∫softmax(hst(z,θ))p(θ|f)θp(y|z,P_f)≈ (h_st(z;θ))p(θ|P_f)dθ (10) To derive a source-calibrated uncertainty weight wiw_i for each target sample xix_i, the predictive posterior is approximated using Monte Carlo (MC) integration. Specifically, α parameter samples θj∼(θ|θMAP,H−1) _j (θ| _MAP,H^-1) are drawn to compute the mean predictive probability P¯(xi) P(x_i). The final weight is defined as the exponentiated negative entropy of this mean prediction: wi=exp(−Entropy(P¯(xi)))w_i= (-Entropy( P(x_i))) (11) This weighting mechanism down-weights target samples that exhibit high predictive uncertainty, thereby mitigating negative adaptation on out-of-distribution samples. The uncertainty weights wiw_i are applied to two loss components. First, the Consistency Loss (ℒConsistencyL_Consistency) quantifies the agreement between the outputs of the student and teacher models: ℒConsistency= _Consistency= 1N∑i=1Nwi⋅SCE(Mst(xi),Mte(xi)) 1N _i=1^Nw_i·SCE(M_st(x_i),M_te(x_i)) (12) +1N∑i=1Nwi⋅SCE(Mst(x~i),Mte(xi)), + 1N _i=1^Nw_i·SCE(M_st( x_i),M_te(x_i)), where x~i x_i is an augmented version of xix_i, and SCE(p,q)=12(CE(p,q)+CE(q,p))SCE(p,q)= 12(CE(p,q)+CE(q,p)) is the symmetric cross-entropy. Second, the Contrastive Loss (ℒContrastiveL_Contrastive) incorporates uncertainty weighting: ℒContrastive=1N∑i=1Nwi⋅ℒcontrastive(xi)L_Contrastive= 1N _i=1^Nw_i·L_contrastive(x_i) (13) I-D Overall Optimization Objective The total loss for DIPTTA comprises three components: the Consistency Loss (ℒConsistencyL_Consistency), the Contrastive Loss (ℒContrastiveL_Contrastive), and the Replay Loss (ℒReplayL_Replay). The Replay Loss is computed on the feature prototypes driven by the latent prototypes of the source domain to prevent catastrophic forgetting via the replay mechanism: ℒReplay=1N∑i=1NCE(hst(f[y]),y)L_Replay= 1N _i=1^NCE(h_st(P_f[y]),y) (14) The final optimization objective is to minimize the total loss (ℒtotalL_total), a weighted sum of these components: ℒtotal=λ1ℒReplay+λ2ℒContrastive+λ3ℒConsistencyL_total= _1L_Replay+ _2L_Contrastive+ _3L_Consistency (15) where λ1 _1, λ2 _2, and λ3 _3 are hyperparameters controlling the contribution of each loss term. The losses are weighted using λ1=0.5 _1=0.5, λ2=0.25 _2=0.25, and λ3=0.15 _3=0.15 across all experiments (refer to Section IV-F3 and Figure 3 for details on the hyperparameter sensitivity analysis). IV EXPERIMENTS IV-A Datasets and Experiment Details Extensive experiments are conducted to demonstrate the effectiveness of the proposed approach. DIPTTA is evaluated on five benchmark tasks for continual test-time adaptation in image processing: CIFAR-10-C, CIFAR-100-C, ImageNet-C [51], ImageNet-R [52], and the C benchmark. These tasks are designed to assess the robustness of machine learning models to corruptions and disturbances in the input data. TABLE I: Classification error rate (%) on CIFAR10-C, CIFAR100-C, and TinyImageNet-C under TTA. Bold text indicates the best. Method Gauss. Shot Imp. Def. Glass Mot. Zoom Snow Fro. Fog Bri. Con. Ela. Pix. JPEG Mean CIFAR10 to CIFAR10-C Source Only 72.3 65.7 72.9 46.9 54.3 34.8 42.0 25.1 41.3 26.0 9.3 46.7 26.6 58.5 30.3 43.5 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 24.7 22.3 32.4 11.6 32.2 13.0 11.0 16.0 16.1 13.1 7.7 11.2 22.0 17.2 23.6 18.3 EATA [39][ICML 2022]_ [rgb]0.5,0.5,0.5 ~[ICML 2022] 24.3 22.3 32.2 11.3 31.7 12.9 10.8 16.0 16.2 13.4 7.9 11.3 21.5 17.0 23.4 18.1 MEMO [53] 57.2 51.2 56.3 32.3 48.7 27.6 29.6 20.7 29.3 20.7 8.6 30.5 25.1 52.4 27.4 34.4 SAR [9][ICLR 2023]_ [rgb]0.5,0.5,0.5 ~[ICLR 2023] 28.3 26.0 35.7 12.7 34.8 13.9 12.0 17.5 17.6 14.9 8.2 13.0 23.5 19.5 27.2 20.3 COME [54] 24.9 22.1 29.0 16.9 33.1 17.9 15.0 15.6 12.4 12.8 7.9 9.4 21.7 16.0 19.8 18.3 AEA [55] 25.8 22.6 31.6 11.8 30.2 12.5 10.2 14.5 14.1 12.4 7.6 10.8 19.1 14.6 20.1 17.2 DIPTTA (Ours) 23.9 20.9 28.0 11.1 27.3 12.2 10.6 14.7 14.3 11.8 8.4 10.6 18.9 14.0 19.5 16.4 CIFAR100 to CIFAR100-C Source Only 73.0 68.0 39.4 29.3 54.1 30.8 28.8 39.5 45.8 50.3 29.5 55.1 37.2 74.7 41.2 46.4 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 37.3 34.8 34.5 25.0 37.4 27.5 25.1 30.4 32.0 33.8 24.1 28.1 32.9 28.4 36.9 31.2 EATA [39][ICML 2022]_ [rgb]0.5,0.5,0.5 ~[ICML 2022] 37.2 35.1 34.5 25.0 37.0 27.5 25.2 30.4 31.1 34.6 23.9 27.7 32.7 28.4 36.5 31.1 MEMO [53] 57.5 53.5 35.9 28.7 45.5 29.5 29.3 33.9 35.7 46.5 26.6 38.3 36.9 52.2 38.8 39.2 SAR [9][ICLR 2023]_ [rgb]0.5,0.5,0.5 ~[ICLR 2023] 40.5 37.9 38.7 26.6 39.9 28.9 27.0 33.4 33.2 38.7 25.5 29.4 34.2 31.4 39.4 33.6 COME [54][ICLR 2025]_ [rgb]0.5,0.5,0.5 ~[ICLR 2025] 37.3 34.8 34.4 25.0 37.4 27.6 25.0 30.4 31.9 33.7 24.0 28.1 32.9 28.3 36.9 31.1 AEA [55][ICLR 2025]_ [rgb]0.5,0.5,0.5 ~[ICLR 2025] 36.9 34.4 33.8 24.8 37.4 27.2 24.6 30.0 31.1 33.7 23.5 28.1 32.7 28.3 36.4 30.9 DIPTTA (Ours) 38.8 32.5 37.5 25.0 40.9 24.4 23.7 30.3 33.5 32.8 19.1 21.6 24.7 31.9 29.2 30.4 ImageNet to TinyImageNet-C Source Only 96.6 95.1 97.2 92.5 92.2 77.8 78.5 81.9 78.1 89.5 77.8 98.3 69.5 71.9 55.6 83.5 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 66.5 64.3 72.9 63.2 75.1 55.5 54.8 63.6 61.9 67.6 57.6 86.2 56.6 51.5 52.5 63.3 EATA [39][ICML 2022]_ [rgb]0.5,0.5,0.5 ~[ICML 2022] 65.1 63.6 69.7 62.3 74.2 55.3 54.1 61.3 61.0 63.9 55.4 90.8 56.1 51.1 52.2 62.4 SAR [9][ICLR 2023]_ [rgb]0.5,0.5,0.5 ~[ICLR 2023] 67.2 65.2 73.6 63.9 75.9 55.8 55.2 64.1 62.6 68.4 58.0 86.0 57.1 51.8 52.9 63.8 COME [54][ICLR 2025]_ [rgb]0.5,0.5,0.5 ~[ICLR 2025] 66.1 64.6 72.3 63.2 75.1 55.8 54.9 63.1 61.2 65.9 56.9 85.3 56.9 51.8 52.9 63.1 AEA [55][ICLR 2025]_ [rgb]0.5,0.5,0.5 ~[ICLR 2025] 63.7 62.1 67.9 60.9 72.8 53.8 52.9 60.1 58.9 61.9 54.2 90.3 54.5 49.8 51.1 61.0 DIPTTA (Ours) 63.4 62.8 66.8 65.4 68.8 53.3 51.5 58.3 58.2 58.6 53.9 90.1 56.2 47.9 50.1 60.1 In all experiments, the TTA setup is strictly adhered to, wherein no source data is accessed. All models are evaluated online, based on a maximum corruption severity level of five. Similar to CoTTA, standard pre-trained WideResNet [56], ResNeXt-29 [57], and ResNet-50 [58] are employed as the source models on a single RTX3090 (24GB) GPU for CIFAR10-C, CIFAR100-C, ImageNet-C, TinyImageNet-C, ImageNet-R, and the C benchmark. All results are evaluated with a corruption severity level of 5 in an online manner. For test time adaptation, the learning rate is set to 0.00025/0.001 for ResNet50/ViT-Base experiments. SGD [59] is utilized as the optimizer, with a momentum of 0.9 and a batch size of 64. IV-B Comparison with SOTA methods for TTA Table I presents the comparative results of DIPTTA with state-of-the-art methods, including entropy minimization and test-time batch normalization under the setting of TTA. IV-B1 Comparison on CIFAR10 to CIFAR10-C The experimental results on the CIFAR10-C benchmark demonstrate the superiority of the proposed DIPTTA method. As shown in the table, DIPTTA achieves the lowest mean classification error of 16.4%, outperforming the runner-up method AEA (17.2%) by 0.8% and significantly reducing the error compared to the Source Only baseline (43.5%). Specifically, DIPTTA attains the best performance in 10 out of 15 corruption types, showing particular robustness in categories such as Gaussian noise, Shot noise, and Impulse noise, which clearly demonstrates the effectiveness of DIPTTA in addressing continual domain shifts and its advantage over the state-of-the-art methods. IV-B2 Comparison on CIFAR100 to CIFAR100-C On the more challenging CIFAR100-C dataset, DIPTTA continues to demonstrate robust adaptation capabilities. It achieves a state-of-the-art mean error rate of 30.4%, surpassing competitive baselines such as AEA (30.9%) and EATA (31.1%). The proposed method shows significant improvements in severe corruption scenarios, particularly excelling in Contrast (21.6%) and Elastic transform (24.7%), where it outperforms the second-best results by substantial margins. IV-B3 Comparison on ImageNet to TinyImageNet-C For the large-scale TinyImageNet-C benchmark, DIPTTA maintains its leadership with a mean error of 60.1%. This represents a substantial improvement over the Source Only baseline of 83.5% and outperforms the previous method, AEA (61.0%). The results indicate that DIPTTA scales effectively to complex datasets, achieving the lowest error rates across a diverse range of corruptions, including Glass blur, Snow, and Pixelate. IV-C Comparison with SOTA methods for CTTA The proposed DIPTTA is compared with various CTTA benchmark methods, including entropy regularization, unsupervised contrastive learning, self-training, and pseudo-label filtering. The results are shown in Table I for CIFAR10-C, CIFAR100-C, ImageNet-C, and in Table I for the C benchmark, respectively. IV-C1 Comparison on CIFAR10 to CIFAR10-C On the CIFAR10-C benchmark, directly testing the Source-only model on the target domains yields a high average error of 43.5%. While subsequent methods like Tent, which aids in sequential adaptation, and CoTTA, which enhances pseudo-label quality, improve upon this baseline, the proposed DIPTTA method establishes a new state-of-the-art. It achieves the lowest mean error rate of 13.9%, significantly outperforming all competing methods, including RMT (16.7%), BeCoTTA (16.3%), and TCA (14.7%). This demonstrates its superior ability to adapt to continual domain shifts in this setting. TABLE I: Classification error rate (%) on CIFAR10-C, CIFAR100-C, and ImageNet-C under CTTA. Bold text indicates the best. Method Gauss. Shot Imp. Def. Glass Mot. Zoom Snow Fro. Fog Bri. Con. Ela. Pix. JPEG Mean CIFAR10 to CIFAR10-C Source Only 72.3 65.7 72.9 46.9 54.3 34.8 42.0 25.1 41.3 26.0 9.3 46.7 26.6 58.5 30.3 43.5 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 24.8 20.6 28.6 14.4 31.1 16.5 14.1 19.1 18.6 18.6 12.2 20.3 25.7 20.8 24.9 20.7 CoTTA [10][CVPR 2022]_ [rgb]0.5,0.5,0.5 ~[CVPR 2022] 24.3 21.3 26.6 11.6 27.6 12.2 10.3 14.8 14.1 12.4 7.6 10.6 18.3 13.4 17.3 16.2 RMT [5][CVPR 2023]_ [rgb]0.5,0.5,0.5 ~[CVPR 2023] 24.0 20.4 25.6 12.6 25.4 14.2 12.2 15.4 15.1 14.1 10.3 13.7 17.1 13.5 16.0 16.7 DSS [60][WACV 2024]_ [rgb]0.5,0.5,0.5 ~[WACV 2024] 24.1 21.3 25.4 11.7 26.9 12.2 10.5 14.5 14.1 12.5 7.8 10.8 18.0 13.1 17.3 16.0 BeCoTTA [61][ICML 2024]_ [rgb]0.5,0.5,0.5 ~[ICML 2024] 22.9 19.1 26.9 10.2 27.5 12.7 10.4 14.7 14.3 12.4 7.2 9.4 20.9 15.2 20.2 16.3 BeCoTTA [61][CVPR 2025]_ [rgb]0.5,0.5,0.5 ~[CVPR 2025] 22.4 19.2 23.0 10.8 23.2 11.6 9.9 13.1 13.1 11.8 7.6 10.7 16.4 12.0 15.5 14.7 DIPTTA (Ours) 21.9 18.2 24.7 10.4 23.8 11.2 9.5 12.1 11.8 10.3 7.1 8.6 14.7 10.4 14.2 13.9 CIFAR100 to CIFAR100-C Source Only 73.0 68.0 39.4 29.3 54.1 30.8 28.8 39.5 45.8 50.3 29.5 55.1 37.2 74.7 41.2 46.4 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 37.2 35.8 41.7 37.9 51.2 48.3 48.5 58.4 63.7 71.1 70.4 82.3 88.0 88.5 90.4 60.9 CoTTA [10][CVPR 2022]_ [rgb]0.5,0.5,0.5 ~[CVPR 2022] 40.1 37.7 39.7 26.9 38.0 27.9 26.4 32.8 31.8 40.3 24.7 26.9 32.5 28.3 33.5 32.5 RMT [5][CVPR 2023]_ [rgb]0.5,0.5,0.5 ~[CVPR 2023] 40.5 36.1 36.3 27.7 33.9 28.5 26.4 29.0 29.0 32.5 25.1 27.4 28.2 26.3 29.3 30.4 DSS [60][WACV 2024]_ [rgb]0.5,0.5,0.5 ~[WACV 2024] 39.7 36.0 37.2 26.3 35.6 27.5 25.2 31.4 30.0 37.8 24.2 26.0 30.0 26.3 31.3 30.9 BeCoTTA [61][ICML 2024]_ [rgb]0.5,0.5,0.5 ~[ICML 2024] 42.1 38.0 42.2 30.2 42.9 31.7 29.8 35.1 33.9 38.5 27.9 32.0 36.7 31.6 39.9 35.5 BeCoTTA [61][CVPR 2025]_ [rgb]0.5,0.5,0.5 ~[CVPR 2025] 38.5 36.0 36.6 25.8 34.6 27.2 25.1 30.5 27.0 30.1 24.1 25.7 27.3 26.6 30.3 29.7 DIPTTA (Ours) 37.4 34.0 34.6 25.1 32.4 25.7 23.5 26.3 25.9 29.1 22.4 23.5 25.7 23.2 27.6 27.8 ImageNet to ImageNet-C Source Only 97.8 97.1 98.2 81.7 89.8 85.2 78.0 83.5 77.1 75.9 41.3 94.5 82.5 79.3 68.6 82.0 Tent [8][ICLR 2021]_ [rgb]0.5,0.5,0.5 ~[ICLR 2021] 81.6 74.6 72.7 77.6 73.8 65.5 55.3 61.6 63.0 51.7 38.2 72.1 50.8 47.4 53.3 62.6 CoTTA [10][CVPR 2022]_ [rgb]0.5,0.5,0.5 ~[CVPR 2022] 84.7 82.1 80.6 81.3 79.0 68.6 57.5 60.3 60.5 48.3 36.6 66.1 47.3 41.2 46.0 62.7 RMT [5][CVPR 2023]_ [rgb]0.5,0.5,0.5 ~[CVPR 2023] 80.2 76.4 74.5 77.1 74.4 66.2 57.6 57.0 59.1 48.0 39.1 60.6 47.3 42.5 43.4 60.2 DSS [60][WACV 2024]_ [rgb]0.5,0.5,0.5 ~[WACV 2024] 82.3 78.4 76.7 81.9 77.8 66.9 60.9 50.8 60.9 47.7 35.4 69.0 47.5 40.9 46.2 62.2 BeCoTTA [61][ICML 2024]_ [rgb]0.5,0.5,0.5 ~[ICML 2024] 84.1 74.3 72.2 77.4 71.9 63.4 55.1 57.2 61.2 50.7 36.4 66.1 49.2 45.6 48.4 60.9 BeCoTTA [61][CVPR 2025]_ [rgb]0.5,0.5,0.5 ~[CVPR 2025] 78.3 71.8 73.5 74.4 73.5 63.3 56.5 56.9 59.4 48.1 39.6 59.6 47.2 42.9 44.7 59.3 DIPTTA (Ours) 76.9 70.3 72.5 73.8 72.7 62.2 56.0 56.1 58.3 47.4 38.8 58.4 46.5 42.0 43.9 58.4 TABLE I: Classification accuracy (%) on C-benchmark, which is a long-sequence task. Bold text indicates the best. Method C-Easy C-Medium C-Hard Average CoTTA [10][CVPR 2022]_ [rgb]0.5,0.5,0.5 ~[CVPR 2022] 14.9±0.8814.9± 0.88 7.7±0.437.7± 0.43 1.1±0.161.1± 0.16 7.9 ETA [39][ICML 2022]_ [rgb]0.5,0.5,0.5 ~[ICML 2022] 41.4±0.9541.4± 0.95 1.1±0.431.1± 0.43 0.2±0.050.2± 0.05 14.2 EATA [39][ICML 2022]_ [rgb]0.5,0.5,0.5 ~[ICML 2022] 48.2±0.6048.2± 0.60 35.4±1.0235.4± 1.02 8.7±0.808.7± 0.80 30.8 RDumb [62][NeurIPS 2023]_ [rgb]0.5,0.5,0.5 ~[NeurIPS 2023] 49.3±0.8849.3± 0.88 38.9±1.4038.9± 1.40 9.6±1.609.6± 1.60 32.6 BeCoTTA [61][CVPR 2025]_ [rgb]0.5,0.5,0.5 ~[CVPR 2025] 49.1±0.3549.1± 0.35 39.5±0.5339.5± 0.53 10.1±0.2210.1± 0.22 32.9 DIPTTA (Ours) 51.2±0.7051.2± 0.70 40.7±0.7240.7± 0.72 9.4±0.579.4± 0.57 33.8 TABLE IV: Classification error rate (%) on ImageNet-R. Results are evaluated in the single-domain adaptation scenario. Bold text indicates the best. We use * to denote episodic adaptation and use ′ to denote the re-implemented. Model Source Only Tent EATA MEMO* CoTTA RMT TEA SAR COME AEA EATA-C′ TCA DIPTTA(Ours) ResNet-50(BN) 62.0 57.7 55.1 58.1 57.6 57.1 57.2 57.3 54.6 54.2 52.9 53.4 52.7 ViT(LN) 47.5 45.8 41.8 42.5 43.6 45.8 39.9 45.0 38.5 37.3 35.8 36.2 35.3 TABLE V: An ablation study of the DFPR module and the SCUE module across four datasets, where (C) indicates the CTTA, (BN) indicates employing the ResNet50 as backbone, and (LN) indicates employing ViT as backbone. Prototype Uncertainty Example Method CIFAR 10-C CIFAR 100-C Tiny ImageNet-C ImageNet -R(BN) ImageNet -R(LN) CIFAR 10-C(C) CIFAR 100-C(C) ImageNet -C(C) None Entropy-based Tent 18.3 31.2 63.3 57.7 43.8 20.7 60.9 62.6 None Entropy-based COME 18.3 31.1 63.1 54.6 38.5 16.7 29.0 58.9 Static Entropy-based RMT(baseline) 18.2 31.0 62.8 57.1 45.8 16.7 30.4 60.2 Static Source-Calibrated Ours(w/o DFPR) 17.2 30.8 61.7 54.0 37.9 15.2 28.1 59.0 Dynamic(DFPR) Entropy-based Ours(w/o SCUE) 16.9 30.9 61.3 53.4 38.7 14.5 28.7 59.6 Dynamic(DFPR) Source-Calibrated Ours 16.4 30.4 60.1 52.7 35.3 13.9 27.8 58.4 IV-C2 Comparison on CIFAR100 to CIFAR100-C For the more challenging CIFAR100-C dataset, the Source-only model starts with a 46.4% error rate. It clearly exposes the critical issue of error accumulation in continual test-time adaptation. For example, the Tent-based method suffers from this problem, resulting in a substantially higher error rate of 60.9% in the long run. Methods like RMT, which employs a symmetric cross-entropy optimization, DSS, which notes the existence of high- and low-quality samples, and BeCoTTA, which improves computational efficiency, all offer incremental progress. On the other hand, the DIPTTA method once again demonstrates dominant performance, reducing the average error to just 27.8%. This result is a significant improvement over all prior art and underscores its effectiveness in preventing model degradation over long, sequential adaptations. IV-C3 Comparison on ImageNet to ImageNet-C On the large-scale ImageNet-C dataset, the source-only model struggles immensely, showing an initial error rate of 82.0%. This benchmark tests the scalability and robustness of adaptation methods under severe corruptions. Although advanced approaches such as CoTTA (62.7%) provide substantial gains, they still fall short of the top performance. The DIPTTA method again proves its robustness, achieving a leading mean error rate of 58.4%. It outperforms the latest methods like TCA (59.3%) and BeCoTTA (60.9%), confirming its effectiveness on complex, large-scale data. IV-C4 Comparison on ImageNet to C To further assess the robustness of the DIPTTA model in long-sequence continual adaptation tasks, an evaluation using the ImageNet to C benchmark was conducted. As detailed in Table I, the DIPTTA method shows highly competitive results. Specifically, DIPTTA achieves leading performance on both C-Easy and C-Medium difficulty levels, with accuracies of 51.2% and 40.7%, respectively. Although TCA performs the best on the C-Hard setting, DIPTTA obtains the best overall performance when considering the three difficulties, registering an average accuracy of 33.8%. IV-D Comparison with SOTA under Natural Domain Shifts As shown in Table IV, experimental evaluation on the ImageNet-R dataset substantiates the superior and consistent performance of the proposed DIPTTA method. On the ResNet-50 architecture, DIPTTA achieves a state-of-the-art error rate of 52.7%, and its advantage is even more pronounced on the Vision Transformer (ViT) architecture, where it reduces the error rate to 35.3%. This consistent top-tier performance validates the potent effectiveness and generalizability of DIPTTA. Its success is attributed to the distilling image prototype, which captures essential and abstract class knowledge rather than just superficial features. This allows the model to effectively distinguish between the core content of an image and its changing artistic style, maintaining strong recognition capabilities while adapting. In contrast, other methods designed primarily for specific algorithmic corruptions struggle to generalize to such complex and varied natural domain shifts. TABLE VI: Time cost(s) per batch on five datasets. We use a batch size of 200 for all results. We use ′ to denote the re-implemented. Dataset Source Only Tent EATA MEMO CoTTA RMT SAR COME AEA EATA-C′ TCA DIPTTA(Ours) CIFAR10-C 0.718 1.809 2.643 62.350 3.893 2.635 3.192 2.930 31.269 2.973 6.186 2.749 CIFAR100-C 0.827 1.992 2.862 73.403 4.127 2.965 3.428 3.357 36.707 3.505 6.539 2.979 ImageNet-C 3.157 7.093 9.561 129.139 17.285 11.773 15.021 16.153 116.250 12.379 26.154 12.235 ImageNet-R(BN) 3.553 7.651 9.664 126.38 18.015 12.036 15.326 16.277 123.135 13.022 27.003 13.173 ImageNet-R(LN) 10.052 22.391 28.095 376.748 73.571 37.268 47.350 50.189 392.267 43.174 98.150 39.258 IV-E Ablation Studies To evaluate the contribution of the two key innovations, a detailed ablation study was performed. Their effectiveness is analyzed by comparing the results with existing methods and by evaluating the standalone impact of each component. IV-E1 Effectiveness of Dynamic Feature Prototype Generation The results from the CTTA Setting highlight the critical need for dynamic prototypes to prevent catastrophic forgetting. For instance, the Tent method, which lacks a knowledge replay mechanism, sees its error rate on CIFAR100-C increase dramatically from 31.2% in the TTA setting to 60.9% under CTTA. More importantly, the DIPTTA method, which uses dynamic prototypes, significantly outperforms RMT, a method that relies on static prototypes, across all datasets. On CIFAR10-C, DIPTTA achieves a 13.9% error rate compared to an error rate of 16.7% for RMT. This confirms that dynamically regenerating prototypes that stay aligned with the evolving model is crucial for preventing the knowledge anchor from becoming obsolete. IV-E2 Advantage of Source-Calibrated Uncertainty The source-calibrated uncertainty mechanism demonstrates clear advantages over conventional entropy-based methods. DIPTTA consistently outperforms methods like Tent and EATA, which are susceptible to overconfident predictions on distribution-shifted data. On the challenging ImageNet-R benchmark, which features natural domain shifts, DIPTTA achieves state-of-the-art results. Furthermore, it accomplishes this while maintaining high computational efficiency, offering a superior trade-off compared to other methods. This performance validates that calibrating uncertainty against stable source knowledge leads to more reliable and robust adaptation. IV-E3 Component-wise Analysis A direct component-wise ablation, detailed in Table V, confirms the individual and combined value of the DFPR and SCUE modules. The baseline model (equivalent to Tent), without either component, records a reference error rate of 18.3% on CIFAR10-C. Integrating only the SCUE module (DIPTTA w/o DFPR) lowers the error rate to 17.2%, confirming its standalone value in mitigating error accumulation. Similarly, incorporating only the DFPR module (DIPTTA w/o SCUE) provides an even greater boost by reducing the error rate to 16.9%. Most importantly, the combination of both DFPR and SCUE in the full DIPTTA model achieves the best result, dropping the CIFAR10-C error rate to 16.4%, which represents a total improvement of 1.9 percentage points over the baseline. This trend remains consistent across other datasets, such as TinyImageNet-C, where the error rate falls from a baseline of 63.3% to 60.1% when both modules are used. The superior performance of the combined system strongly suggests that the two components are complementary and produce a synergistic effect, justifying the inclusion of both in the final model architecture. IV-F Further Empirical Analysis IV-F1 Time Cost Analysis The time cost analysis in Table VI confirms the exceptional computational efficiency of DIPTTA. In sharp contrast to computationally intensive methods like MEMO and AEA, whose memory banks and meta-learning render them impractical for time-sensitive applications, DIPTTA maintains a lean profile comparable to its efficient baseline, RMT. The minimal overhead from the proposed DFPR and SCUE modules is the key takeaway; on ImageNet-C, DIPTTA (12.235s) is only marginally slower than RMT (11.773s). This result validates that DIPTTA successfully achieves a superior trade-off, advancing adaptation performance without the cost of inefficiency. Fig. 2: Sensitivity analysis of batch size, where C indicates the CTTA. IV-F2 Sensitivity Analysis of Batch Size As illustrated in Figure 2, DIPTTA exhibits robustness to batch size, consistently outperforming competitors even at a batch size of 1. On CIFAR benchmarks, it achieves optimal error rates (13.9% and 27.8%) at batch size 200 while maintaining stability across the 50–300 range, contrasting with the volatility of methods like TCA. This stability extends to ImageNet-C, confirming the low sensitivity and practical reliability of DIPTTA. Fig. 3: Loss function hyperparameter sensitivity analysis of the TTA method on three benchmark datasets. IV-F3 Hyperparameter Sensitivity Analysis Figure 3 illustrates the impact of three hyperparameters (λ1λ2,λ3 _1 _2, _3) on the performance of a TTA method across the CIFAR10-C and CIFAR100-C datasets. The color of each point represents the mean error rate (%), where brighter yellow indicates lower error. The optimal points for the datasets were found to be (0.5, 0.25, 0.15) and (0.45, 0.2, 0.2), respectively. Given the proximity of these optima, the single parameter set of (0.5, 0.25, 0.15), marked by the red star, is selected for all experiments. Fig. 4: Visualization of t-SNE for two methods and our proposed DIPTTA on CIFAR10-C(TTA Scenarios). We select a random batch of features from the Gaussian and Fog domains for comparative visualization. IV-F4 t-SNE Analysis t-SNE [63] is utilized to visualize the feature representations of the model. As shown in Fig. 4, the feature vectors (dots) from the DIPTTA model exhibit significantly tighter clustering around the source domain prototypes (stars) compared to other methods on CIFAR10-C. This improved performance stems from two key factors: 1) the employed data distillation technique refines source domain knowledge to create more accurate class prototypes that better represent true feature centroids; and 2) the combined uncertainty-weighted consistency and contrastive losses form a robust constraint that aligns target feature vectors with these optimized prototypes. Consequently, the model learns more compact and separable features, creating clearer decision boundaries. V CONCLUSION In this paper, the critical challenges of error accumulation from noisy pseudo-labels and catastrophic forgetting of source knowledge during test-time adaptation are analyzed. To this end, the DIPTTA framework is proposed, which introduces a Distilling Image Prototype (DIP) as a compact and regenerative source knowledge anchor. The core contribution is twofold. First, the DIP enables dynamic feature prototype replay, which continuously regenerates feature prototypes aligned with the evolving feature space of the model, effectively overcoming the misalignment issue of static prototypes. Second, it facilitates source-calibrated uncertainty estimation, which leverages the stable DIP to debias uncertainty quantification and reliably suppress error propagation. Extensive experiments on multiple benchmarks validate that DIPTTA significantly outperforms state-of-the-art methods. The results demonstrate that the proposed approach provides a robust and effective solution for model adaptation in non-stationary test environments. References [1] J. Liang, R. He, and T. Tan (2025) A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision 133 (1), p. 31–64. Cited by: §I, §I. [2] Z. Wang, Y. Luo, L. Zheng, Z. Chen, S. Wang, and Z. Huang (2025) In search of lost online test-time adaptation: a survey. International Journal of Computer Vision 133 (3), p. 1106–1139. Cited by: §I, §I. [3] L. Chen, Y. Zhang, Y. Song, Y. Shan, and L. Liu (2023) Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 24172–24182. Cited by: §I. [4] B. Hu, J. Liu, K. Zheng, and Z. Zha (2023) Unleashing knowledge potential of source hypothesis for source-free domain adaptation. IEEE Transactions on Multimedia 26, p. 5422–5434. Cited by: §I. [5] M. Döbler, R. A. Marsden, and B. Yang (2023) Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 7704–7714. Cited by: §I, §I-A, §I-A, TABLE I, TABLE I, TABLE I. [6] G. Chakrabarty, M. Sreenivas, and S. Biswas (2023) Santa: source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research. Cited by: §I, §I-A. [7] M. Jang, S. Chung, and H. W. Chung (2022) Test-time adaptation via self-training with nearest neighbor information. In The Eleventh International Conference on Learning Representations, Cited by: §I, §I-A. [8] D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell (2020) Tent: fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, Cited by: §I, §I-A, §I-B, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [9] S. Niu, J. Wu, Y. Zhang, Z. Wen, Y. Chen, P. Zhao, and M. Tan (2023) Towards stable test-time adaptation in dynamic wild world. In The Eleventh International Conference on Learning Representations, Cited by: §I, §I-A, TABLE I, TABLE I, TABLE I. [10] Q. Wang, O. Fink, L. Van Gool, and D. Dai (2022) Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 7201–7211. Cited by: §I, TABLE I, TABLE I, TABLE I, TABLE I. [11] Z. Xiao and C. G. Snoek (2024) Beyond model adaptation at test time: a survey. arXiv preprint arXiv:2411.03687. Cited by: §I. [12] P. Tian, X. Zhou, H. Yu, and S. Xie (2025) High specificity guided cross-domain few-shot segmentation. IEEE Transactions on Multimedia. Cited by: §I. [13] Y. Gal and Z. Ghahramani (2016) Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the International Conference on Machine Learning (ICML), p. 1050–1059. Cited by: §I, §I-B. [14] D. J. C. Mackay (1992) Bayesian methods for adaptive models. California Institute of Technology. Cited by: §I, §I-B. [15] A. Kristiadi, M. Hein, and P. Hennig (2020) Being bayesian, even just a bit, fixes overconfidence in relu networks. In International conference on machine learning, p. 5436–5446. Cited by: §I. [16] E. Daxberger, A. Kristiadi, A. Immer, R. Eschenhagen, M. Bauer, and P. Hennig (2021) Laplace redux-effortless bayesian deep learning. Advances in neural information processing systems 34, p. 20089–20103. Cited by: §I. [17] B. Zhao and H. Bilen (2021) Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, p. 12674–12685. Cited by: §I. [18] L. Yuan, B. Xie, and S. Li (2023) Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15922–15932. Cited by: §I-A. [19] J. Ma (2024) Improved self-training for test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 23701–23710. Cited by: §I-A. [20] S. Goyal, M. Sun, A. Raghunathan, and J. Z. Kolter (2022) Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems 35, p. 6204–6218. Cited by: §I-A. [21] S. Sinha, P. Gehler, F. Locatello, and B. Schiele (2023) Test: test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 2759–2769. Cited by: §I-A. [22] Y. Iwasawa and Y. Matsuo (2021) Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems 34, p. 2427–2440. Cited by: §I-A. [23] M. Boudiaf, R. Mueller, I. Ben Ayed, and L. Bertinetto (2022) Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8344–8353. Cited by: §I-A. [24] S. Wang, D. Zhang, Z. Yan, J. Zhang, and R. Li (2023) Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20050–20060. Cited by: §I-A. [25] S. Jung, J. Lee, N. Kim, A. Shaban, B. Boots, and J. Choo (2023) Cafa: class-aware feature alignment for test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 19060–19071. Cited by: §I-A. [26] A. Tarvainen and H. Valpola (2017) Mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30. Cited by: §I-A. [27] S. Zhang, L. Zhang, and Z. Liu (2025) Test-time adaptation for object detection via dynamic dual teaching. Image and Vision Computing, p. 105740. Cited by: §I-A. [28] Y. Ye, W. Wei, L. Zhang, C. Ding, and Y. Zhang (2025) Domain consistency learning for continual test-time adaptation in image semantic segmentation. Pattern Recognition 165, p. 111585. Cited by: §I-A. [29] D. Chen, D. Wang, T. Darrell, and S. Ebrahimi (2022) Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 295–305. Cited by: §I-A. [30] Z. Ma, Y. Li, Y. Luo, X. Luo, J. Li, C. Chen, X. Hua, and G. Lu (2024) Discrepancy and structure-based contrast for test-time adaptive retrieval. IEEE Transactions on Multimedia 26, p. 8665–8677. Cited by: §I-A. [31] T. Gong, J. Jeong, T. Kim, Y. Kim, J. Shin, and S. Lee (2022) Note: robust continual test-time adaptation against temporal correlation. Advances in Neural Information Processing Systems 35, p. 27253–27266. Cited by: §I-A. [32] D. Sójka, S. Cygert, B. Twardowski, and T. Trzciński (2023) Ar-tta: a simple method for real-world continual test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 3491–3495. Cited by: §I-A. [33] F. F. Niloy, S. M. Ahmed, D. S. Raychaudhuri, S. Oymak, and A. K. Roy-Chowdhury (2024) Effective restoration of source knowledge in continual test time adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 2091–2100. Cited by: §I-A. [34] G. Wang, C. Ding, W. Tan, and M. Tan (2025) Decoupled prototype learning for reliable test-time adaptation. IEEE Transactions on Multimedia. Cited by: §I-A. [35] J. Geng, Z. Chen, Y. Wang, H. Woisetschlaeger, S. Schimmler, R. Mayer, Z. Zhao, and C. Rong (2023) A survey on dataset distillation: approaches, applications and future directions. In IJCAI, Cited by: §I-A. [36] A. Carta, A. Cossu, V. Lomonaco, and D. Bacciu (2022) Distilled replay: overcoming forgetting through synthetic samples. In Continual Semi-Supervised Learning: First International Workshop, CSSL 2021, Virtual Event, August 19-20, 2021, Revised Selected Papers, Vol. 13418, p. 104. Cited by: §I-A. [37] H. Jin, S. Liu, C. Cong, Q. Feng, Y. Liu, L. Huang, and Y. Hu (2025) Fedwsidd: federated whole slide image classification via dataset distillation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, p. 178–188. Cited by: §I-A. [38] X. Dong, L. Wang, X. Lv, X. Zhang, H. Zhang, B. Pu, Z. Gao, I. Y. Liao, and Z. Jin (2025) CertainTTA: estimating uncertainty for test-time adaptation on medical image segmentation. Information Fusion, p. 103300. Cited by: §I-B. [39] S. Niu, J. Wu, Y. Zhang, Y. Chen, S. Zheng, P. Zhao, and M. Tan (2022) Efficient test-time model adaptation without forgetting. In International conference on machine learning, p. 16888–16905. Cited by: §I-B, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [40] M. Tan, G. Chen, J. Wu, Y. Zhang, Y. Chen, P. Zhao, and S. Niu (2025) Uncertainty-calibrated test-time model adaptation without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence. Cited by: §I-B. [41] J. Liang, D. Hu, and J. Feng (2020) Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), p. 6028–6039. Cited by: §I-B. [42] R. Rahaman et al. (2021) Uncertainty quantification and deep ensembles. Advances in neural information processing systems 34, p. 20063–20075. Cited by: §I-B. [43] T. Abe, E. K. Buchanan, G. Pleiss, R. Zemel, and J. P. Cunningham (2022) Deep ensembles work, but are they necessary?. Advances in Neural Information Processing Systems 35, p. 33646–33660. Cited by: §I-B. [44] J. Denker and Y. LeCun (1990) Transforming neural-net output levels to probability distributions. Advances in neural information processing systems 3. Cited by: §I-B. [45] D. Hendrycks and K. Gimpel (2017) A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, Cited by: §I-B. [46] W. Liu, X. Wang, J. Owens, and Y. Li (2020) Energy-based out-of-distribution detection. Advances in neural information processing systems 33, p. 21464–21475. Cited by: §I-B. [47] P. C. Trimmer, A. I. Houston, J. A. Marshall, M. T. Mendl, E. S. Paul, and J. M. McNamara (2011) Decision-making under uncertainty: biases and bayesians. Animal cognition 14 (4), p. 465–476. Cited by: §I-B. [48] M. Tan, P. Chen, H. Zhi, J. Mai, B. Rosman, D. Ji, and R. Zeng (2025) Source-free elastic model adaptation for vision-and-language navigation. IEEE Transactions on Multimedia. Cited by: §I-B. [49] A. Kumar, P. S. Liang, and T. Ma (2019) Verified uncertainty calibration. Advances in neural information processing systems 32. Cited by: §I-C. [50] H. Ritter, A. Botev, and D. Barber (2018) A scalable laplace approximation for neural networks. In 6th international conference on learning representations, ICLR 2018-conference track proceedings, Vol. 6. Cited by: §I-C. [51] A. Krizhevsky G. Hinton et al. (2009) Learning multiple layers of features from tiny images. Cited by: §IV-A. [52] D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guo, et al. (2021) The many faces of robustness: a critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, p. 8340–8349. Cited by: §IV-A. [53] M. Zhang, S. Levine, and C. Finn (2022) Memo: test time robustness via adaptation and augmentation. Advances in neural information processing systems 35, p. 38629–38642. Cited by: TABLE I, TABLE I. [54] Q. Zhang, Y. Bian, X. Kong, P. Zhao, and C. Zhang (2025) COME: test-time adaptation by conservatively minimizing entropy. In International Conference on Learning Representations, Cited by: TABLE I, TABLE I, TABLE I. [55] W. Choi, D. Kim, J. Park, J. Lee, Y. Park, D. Han, and J. Moon (2025) Adaptive energy alignment for accelerating test-time adaptation. In The Thirteenth International Conference on Learning Representations, Cited by: TABLE I, TABLE I, TABLE I. [56] S. Zagoruyko and N. Komodakis (2016) Wide residual networks. In British Machine Vision Conference 2016, Cited by: §IV-A. [57] D. Yin, R. Gontijo Lopes, J. Shlens, E. D. Cubuk, and J. Gilmer (2019) A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems 32. Cited by: §IV-A. [58] F. Croce, M. Andriushchenko, V. Sehwag, E. Debenedetti, N. Flammarion, M. Chiang, P. Mittal, and M. Hein (2020) RobustBench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), Cited by: §IV-A. [59] L. Bottou (2010) Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT, p. 177–186. Cited by: §IV-A. [60] Y. Wang, J. Hong, A. Cheraghian, S. Rahman, D. Ahmedt-Aristizabal, L. Petersson, and M. Harandi (2024) Continual test-time domain adaptation via dynamic sample selection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 1701–1710. Cited by: TABLE I, TABLE I, TABLE I. [61] D. Lee, J. Yoon, and S. J. Hwang (2024) BECoTTA: input-dependent online blending of experts for continual test-time adaptation. In International Conference on Machine Learning, p. 27072–27093. Cited by: TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I. [62] O. Press, S. Schneider, M. Kümmerer, and M. Bethge (2023) Rdumb: a simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems 36, p. 39915–39935. Cited by: TABLE I. [63] L. Van der Maaten and G. Hinton (2008) Visualizing data using t-sne.. Journal of machine learning research 9 (11). Cited by: §IV-F4.