Research map / Label-efficient and robust learning / Domain adaptation and generalization
StyDeSty: Min-Max Stylization and Destylization for Single Domain Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StyDeSty, a min-max game between a stylization module and an AdaIN-based destylization module that maps stylized and source features to a shared latent domain, with NAS choosing the destylization layer position, for single domain generalization.
Problem
Single domain generalization: learning a model that generalizes to unseen domains when only one source domain is available; prior augmentation methods ignore alignment between source and augmented domains.
Approach
Pairs a learnable stylization module (generates stylized samples) with a destylization module (AdaIN layer plus alignment loss that maps features to a shared latent domain), trained adversarially; a Gumbel-softmax NAS selects where the AdaIN destylization layer sits in the backbone.
Key results
- On PACS with photo as the training domain (ResNet-18), StyDeSty reaches 63.25% average accuracy vs 49.06% for L2D, a 13.44-point gain over the best prior method.
- On Digits (train on MNIST, LeNet-5), StyDeSty averages 78.82% vs 74.46% for L2D.
- On CIFAR-10-C, StyDeSty averages 85.35% vs 85.25% for L2D under the leave-one-domain-out PACS setting; on CIFAR-10-C severity 5 it improves over others by 0.94–8.67 points per corruption category.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use StyDeSty instead of L2D | single-domain generalization on Digits · Digits (MNIST source; Avg.) | accuracy: 74.46 -> 78.82 | Helped |
| use StyDeSty instead of L2D | single-domain generalization on PACS · PACS (Photo source; Avg.) | accuracy: 49.06 -> 63.25 | Helped |
| use StyDeSty instead of v-L2D | monocular depth estimation · KITTI (vKITTI source) | Abs Rel: 0.211 -> 0.197 | Helped |
| use StyDeSty instead of L2D | single-domain generalization on DomainNet · DomainNet (Photo source; Avg.) | accuracy: 25.16 -> 27.18 | Helped |
| add the stylization module (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 48.53 -> 63.25 | Helped |
| replace AutoAug with StyDeSty stylization (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 55.13 -> 63.25 | Helped |
| replace a DCGAN generator with StyDeSty stylization (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 56.68 -> 63.25 | Helped |
| add the destylization AdaIN layer (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 50.81 -> 63.25 | Helped |
| use task-aware destylization instead of separate pretrained style transfer (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 60.74 -> 63.25 | Helped |
| add the feature alignment loss (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 59.47 -> 63.25 | Helped |
| add the perceptual alignment term (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 56.38 -> 63.25 | Helped |
| train the task head separately rather than jointly end-to-end (ablation) | StyDeSty · PACS (Photo source; Avg.) | accuracy: 56.63 -> 63.25 | Helped |
Surprising findings
- Adding the destylization plug-in to prior single-DG methods (e.g., baselines in Table 6) improves them, and the combined variant beats the default StyDeSty on one setting.
Limitations
- NAS convergence becomes difficult when many candidate positions (>10) are considered, per the authors.
- Only a single destylization layer is used; multiple destylization layers are left to future work.
- Performance is sensitive to the destylization layer location and requires a NAS stage.
Experimental setup
Datasets: MNIST SVHN MNIST-M SYN USPS CIFAR-10 CIFAR-10-C PACS vKITTI KITTI DomainNet
Benchmarks: Digits CIFAR-10-C PACS KITTI DomainNet
Metrics: classification accuracy Abs Rel Squa Rel RMSE RMSE_log delta<1.25 accuracy
Baselines: Source Only JiGen RSC MMLD ADA M-ADA ME-ADA MixStyle L2D
Builds on: L2D adversarial data augmentation (ADA/M-ADA/ME-ADA) MixStyle AdaIN
Closest papers by meaning
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.86)
- Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.86)
- Adversarial Domain Prompt Tuning and Generation for Single Domain Generalization (CVPR 2025 · similarity 0.86)
- AdvST: Revisiting Data Augmentations for Single Domain Generalization (AAAI 2024 · similarity 0.86)
- Towards domain-invariant Self-Supervised Learning with Batch Styles Standardization (ICLR 2024 · similarity 0.85)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.