---
kind: concept
slug: domain-generalization
title: Domain generalization
summary: >-
  Train on one or more source domains and predict on an unseen target domain that is never used for training or model selection. Methods change the training data, the learned features, or the training strategy. Trade-off: most families add memory or compute, and the survey's benchmark shows small gains over plain ERM.
tags: [transfer, out-of-distribution, survey, graphs, segmentation, vision-language]
papers: ["2103.03097", "2103.15597", "2104.09937", "2202.02466", "2311.16445"]
related: [domain-adaptation, meta-learning, data-augmentation, ensembles, causal-inference, graph-neural-network, semantic-segmentation, domain-randomization, invariant-learning, zero-shot-learning, vision-language-pretraining, contrastive-learning]
status: draft
updated_at: 2026-11-01
---

## Overview

A model is often tested on data that differs from its training data, and its
accuracy drops when the difference is large. Domain generalization (DG) is the
setting where training uses related source domains, none of them the target,
and the model must work on a new domain it has never seen
[arxiv:2103.03097 §I] [arxiv:2103.03097 §II.A]. Most methods need several
sources, but some normalization-based methods work with one source
[arxiv:2103.15597 §1]. A text-side method on a frozen vision-language model
uses no source images at all: its training data is template prompts and class
names [arxiv:2311.16445 §4.2].

The shared invariant is that the target domain is unavailable during training,
not even unlabelled. This is the main difference from domain adaptation, which
may use unlabelled target data [arxiv:2103.03097 §III.A]. The sources and the
target are assumed to be related but not identical, with the same label space
[arxiv:2103.03097 §II.A].

Some data has no domain labels at all. For node-level prediction on one
graph, with no environment for each node, the method has to create its own
training environments. [arxiv:2202.02466 §2.1] [arxiv:2202.02466 §3.2]
describes one way to do this, by editing the graph (see Data manipulation).

The objective is to minimise the risk on the unseen target. Since that risk
cannot be measured during training, every method is a proxy for it
[arxiv:2103.03097 §II.A] [arxiv:2103.03097 eq. 1].

## Variants

The survey sorts methods into three groups. The paper's own taxonomy is in its
§IV; the names below are the groups the page uses. Two methods on this page
are not among the survey's named methods: RobustNet [arxiv:2103.15597] is
placed under Representation learning, Fish [arxiv:2104.09937] under
Learning strategy, and CLAP [arxiv:2311.16445] under Representation learning.
CLAP's training data is text, so its augmentation is in the spirit of data
manipulation, but its learned component is a feature map.

### Data manipulation

Add modified copies of the inputs to the training set, so the model sees more
variation than the sources contain [arxiv:2103.03097 §IV.A].

- *Augmentation.* Standard transforms such as flips, rotations, scaling,
  cropping and noise, plus domain randomisation that varies object placement,
  texture, illumination and camera view [arxiv:2103.03097 §IV.A.1]. Card:
  [Generalizing to Unseen Domains](/papers/2103.03097).
- *Generation.* Samples from VAEs or GANs, and Mixup, which interpolates two
  samples and their labels. Mixup can run in input space or feature space
  [arxiv:2103.03097 §IV.A.2].
- *Adversarial augmentation.* Perturbations chosen by an optimisation goal,
  for example CrossGrad and iterative fictitious-target augmentation. These
  need adversarial training and are hard to optimise [arxiv:2103.03097 §IV.A.1].
  A graph variant learns the edits. Editors add and delete edges of one
  graph to create K virtual environments, and the GNN minimises the mean plus
  the variance of its risks on them. The editors are trained with REINFORCE
  and the variance as reward [arxiv:2202.02466 §3.2] [arxiv:2202.02466 app. A].
  The paper calls this EERM. Its single-level form, with known environments,
  is the objective REx uses [arxiv:2202.02466 §3.2]. Card: [Handling Distribution Shifts on Graphs](/papers/2202.02466).

### Representation learning

Learn a feature map g so that one classifier f works on all sources and the
target. The general form is a task loss plus a regulariser on g
[arxiv:2103.03097 §IV.B].

- *Domain-invariant representations.* Kernel methods (DICA, SCA), domain
  adversarial training (DANN and class-conditional variants), explicit
  alignment with MMD, second-order statistics or Wasserstein distance, and
  feature normalisation (IN, BIN, SNR). The survey says normalisation can be
  combined with the other methods [arxiv:2103.03097 §IV.B.1].
- *Selective whitening.* Penalise only the channel-covariance entries of a
  normalisation layer that move under colour and blur change, and leave the
  others free. RobustNet's instance selective whitening (ISW) does this in
  three instance-normalisation layers. It picks the entries after a warm-up,
  with k-means on their variance [arxiv:2103.15597 §4.3] [arxiv:2103.15597 §4.4].
  Full whitening removes both style and content correlations, and its exact
  form is slow and blocks gradients [arxiv:2103.15597 §3]. The method trains
  with one source domain [arxiv:2103.15597 §1].
- *Invariant risk minimisation.* Learn a representation whose best classifier
  is the same in every domain. The survey gives a gradient-norm surrogate and
  names its linear assumption and failure cases [arxiv:2103.03097 §IV.B.1].
- *Feature disentanglement.* Split features into a domain-shared part and a
  domain-specific part, with a reconstruction loss so that information is not
  lost [arxiv:2103.03097 §IV.B.2]. Families: multi-component analysis, generative
  modelling (DIVA, UFDN) and causality-inspired methods. Related page:
  [causal-inference](/concepts/causal-inference).
- *Text-side style removal.* For a frozen CLIP-style model, a small residual
  MLP is trained on text features only. Each training prompt is paired with a
  copy in which style words are dropped or reordered, and a contrastive loss
  pulls the pair together and pulls the copy toward the bare class name. The
  MLP is then reused on image features at inference, with a scalar weight on
  its residual branch [arxiv:2311.16445 §4.1.2] [arxiv:2311.16445 §4.2]. The
  paper argues that text is easier to edit for style than images
  [arxiv:2311.16445 §4.2]. Card: [CLAP: Isolating Content from Style through
  Contrastive Learning with Augmented Prompts](/papers/2311.16445). Related
  pages: [zero-shot-learning](/concepts/zero-shot-learning),
  [contrastive-learning](/concepts/contrastive-learning).

### Learning strategy

Change how training runs, without changing the data or the feature map
[arxiv:2103.03097 §IV.C].

- *Ensemble learning.* One classifier per source domain or expert, combined
  with learned weights (Mancini et al., DAEL). Costs more memory and compute
  [arxiv:2103.03097 §IV.C.1]. Related page: [ensembles](/concepts/ensembles).
- *Meta-learning.* Split the sources into meta-train and meta-test sets to
  simulate a shift between domains during training. Examples: MLDG, MetaReg,
  MetaVIB [arxiv:2103.03097 §IV.C.2]. Related page:
  [meta-learning](/concepts/meta-learning).
- *Gradient operations.* RSC removes dominant features during training.
  Gradient inner-product maximisation and gradient-matching methods such as
  Fishr align per-domain gradients [arxiv:2103.03097 §IV.C.3]. Fish is a
  first-order way to reach a similar objective. Each meta-step copies the
  weights, takes one gradient step on the copy for each source domain in random
  order, and then moves the weights part of the way toward the copy. No second
  derivatives are needed [arxiv:2104.09937 §3.4] [arxiv:2104.09937 alg. 1].
  Its inner loop has the shape of Reptile's, but it takes one batch from each
  domain instead of several from one task [arxiv:2104.09937 app. A.1]. Card:
  [Gradient Matching for Domain Generalization](/papers/2104.09937). Related
  page: [invariant-learning](/concepts/invariant-learning).
- *Distributionally robust optimisation.* Optimise the worst-case training
  domain, for example GroupDRO, VREx and AdaRNN [arxiv:2103.03097 §IV.C.4].
- *Self-supervised learning.* Jigsaw pretext tasks and contrastive objectives,
  which can be added to other methods. Costs extra compute
  [arxiv:2103.03097 §IV.C.5].
- *Other.* Metric learning, random forests, calibration, lottery-ticket
  substructures, and weight averaging for flatter minima (SWAD)
  [arxiv:2103.03097 §IV.C.6].

## When to use which

The survey does not give a single best method. Its benchmark uses one
backbone and two image datasets, so the results below apply only to that
setting [arxiv:2103.03097 §VII.C]. RobustNet's results cover urban-scene
segmentation only [arxiv:2103.15597 §5.1.1].

| Situation | Use | Why |
|---|---|---|
| The known variation can be written as input transforms (flips, crops, colour, noise) | Standard augmentation from the data manipulation family | Adds variation without changing the objective [arxiv:2103.03097 §IV.A.1] |
| Memory or compute budget is tight | Avoid ensembles and self-supervised add-ons | The survey names their extra memory and compute cost [arxiv:2103.03097 §IV.C.1] [arxiv:2103.03097 §IV.C.5] |
| Sources are known to differ in a way you can simulate | Meta-learning over source splits | Meta-learning simulates a shift between domains during training [arxiv:2103.03097 §IV.C.2] |
| Labels for each training domain are available | Group-based robust optimisation | GroupDRO works on the worst training domain, and the survey says its setup needs domain labels [arxiv:2103.03097 §IV.C.4] |
| Only one labelled source domain, dense prediction, and the inference cost must stay the same | Selective whitening (ISW) in instance-normalised layers | The normalisation route needs one source [arxiv:2103.15597 §1] [arxiv:2103.15597 §2]. The paper reports no inference cost and the same FLOPs [arxiv:2103.15597 tab. 6]. Gains are shown for DeepLabV3+ on urban scenes only [arxiv:2103.15597 tab. 1] |
| Several labelled sources, a gradient-alignment update is wanted, and extra gradient steps per meta-step are affordable | Fish, a first-order inter-domain gradient matching step | Each meta-step takes one gradient step per source domain on a copy of the weights [arxiv:2104.09937 alg. 1]. On the Wilds table Fish is best on four of six datasets. Its CivilComments worst-group accuracy is 72.8 against 56.0 for ERM [arxiv:2104.09937 tab. 3] |
| Very many domains, such as iWildCam (324) or Amazon (7,676) | Plain ERM first, then test Fish against it | Fish is below ERM on iWildCam macro F1 and level on Amazon, and the paper ties this to the domain count [arxiv:2104.09937 tab. 3] [arxiv:2104.09937 §5] |
| You want a baseline first | Plain ERM on pooled sources | On the survey's benchmark, ERM averages 81.2 on PACS and 62.9 on Office-Home, and the best method is 0.8 points above it on each [arxiv:2103.03097 §VII.C] |
| One graph, no environment labels, and the shift is between graph regions or over time | Learned edge edits that create virtual environments (EERM) | The paper reports gains over ERM on Twitch, Facebook-100, Elliptic and OGB-Arxiv, all author-reported [arxiv:2202.02466 §5] [arxiv:2202.02466 tab. 2] [arxiv:2202.02466 tab. 3]. The reference code keeps a dense matrix per editor, which is very large at OGB-Arxiv size (see Pitfalls) |
| A frozen CLIP-style model is used zero-shot or few-shot, no source images are available, and the shift is in prompt wording or style | Text-side style removal (CLAP) | Trained on text only, from template prompts and class names [arxiv:2311.16445 §4.2]. Author-reported zero-shot average over the four datasets is 79.9 against 77.4 for CLIP, and 76.1 against 70.4 on the noised prompt [arxiv:2311.16445 tab. 3]. On DomainNet, training takes 47 minutes against 3.3 hours for the image-generation variant, on one RTX 3090 [arxiv:2311.16445 §5.1] |
| Choosing a model without target data | Training-domain validation | The survey calls it the most common protocol; test-domain validation would use target data [arxiv:2103.03097 §VII.B] |

Benchmark averages from the survey's table, one run per cell:

- PACS: RSC and CORAL both average 82.0; ERM 81.2; DANN 81.4; ANDMask 79.9
  [arxiv:2103.03097 §VII.C].
- Office-Home: RSC averages 63.7; CORAL 63.3; ERM 62.9; ANDMask 61.6
  [arxiv:2103.03097 §VII.C].

RobustNet's table, author-reported and mIoU in percent. Each value is an
average over three runs [arxiv:2103.15597 §5.2.1]. It is trained on GTAV with
DeepLabV3+ and ResNet-50, and scored on Cityscapes (C), BDD-100K (B) and
Mapillary (M) [arxiv:2103.15597 tab. 1]:

- ISW: C 36.58, B 35.20, M 40.33. Baseline: C 28.95, B 25.14, M 28.18.
- IBN-Net (re-implemented by the authors): C 33.85, B 32.30, M 37.75.

Fish's DomainBed table, author-reported test accuracy averaged over seven
datasets, 5 trials, training-domain validation [arxiv:2104.09937 tab. 4]:

- Coral 67.2; Fish 67.1; Mixup 67.1; ERM 67.0.

Fish's main Wilds table, test set, author-reported [arxiv:2104.09937 tab. 3]:

- CivilComments worst-group accuracy: Fish 72.8; ERM 56.0.
- Camelyon17 average accuracy: Fish 74.7; ERM 70.3.

CLAP's zero-shot table, ViT-B/16 CLIP, averaged over PACS, VLCS, OfficeHome
and DomainNet, author-reported [arxiv:2311.16445 tab. 3]. Bare class-name
prompt: CLIP 77.4, Im.Aug 76.1, CLAP 79.9. Noised prompt: CLIP 70.4, Im.Aug
69.9, CLAP 76.1. The baseline is zero-shot CLIP, not a model trained on the
sources, so these rows are not comparable with the ERM rows above.

These numbers come from different papers and settings: the survey's image
benchmark, RobustNet's segmentation benchmark, Fish's DomainBed and Wilds runs,
and CLAP's zero-shot runs. The graph numbers below come from the EERM paper,
which uses its own setting. All are the authors' own runs:

- OGB-Arxiv, inductive, GPRGNN, 2014–2016 test interval: ERM 47.25 ± 0.55,
  EERM 49.88 ± 0.49 [arxiv:2202.02466 tab. 3]. For GraphSAGE on the same
  interval, EERM is 41.55 against 42.09 for ERM [arxiv:2202.02466 tab. 3].

Differences of half a point or less are within what one seed can move. The
paper's own text calls the RSC gains unremarkable [arxiv:2103.03097 §VII.C].
On DomainBed, Fish's average is 0.1 points above ERM [arxiv:2104.09937 tab. 4].

## Implementation sketch

The common core is pooled training over the source domains, plus a
term that each method adds to the loss. The paper's general form is a task
loss plus a weighted regulariser on the feature map [arxiv:2103.03097 §IV.B]:

```pseudocode
# Common core of DG training  [arxiv:2103.03097 §IV.B]
# Input:  minibatches from each source domain, weight lambda
# Output: scalar loss for one step

def dg_step(source_batches, lam):
    # Pool the batches of all source domains  [arxiv:2103.03097 §II.A]
    x, y, d = pool(source_batches)
    # Encode inputs with the shared feature map g  [arxiv:2103.03097 §IV.B]
    z = g(x)
    # Task loss on the shared classifier f  [arxiv:2103.03097 §IV.B]
    loss = cross_entropy(f(z), y)
    # Add the method's regulariser: alignment, invariance penalty,
    # gradient matching, a disentanglement term, or a masked covariance
    # penalty  [arxiv:2103.03097 §IV.B.1-2] [arxiv:2103.15597 eq. 18]
    loss = loss + lam * regulariser(z, y, d)
    return loss
```

Where the variants differ: the regulariser (alignment, invariance penalty,
reconstruction, or a covariance penalty with a fixed mask), the batch sampling
(pooled, per-domain, or worst-domain), the update rule (Fish replaces the single
step with an inner loop over domains on a copy of the weights
[arxiv:2104.09937 alg. 1]), and extra training stages for meta-learning
[arxiv:2103.03097 §IV.B] [arxiv:2103.03097 §IV.C]. A selective
whitening method computes its mask once, after a warm-up stage, and then
keeps it fixed [arxiv:2103.15597 §4.3]. With no domain labels, the "sources"
are edited copies of one graph, and the loss is taken over the copies' risks
[arxiv:2202.02466 §3.2]. A text-only variant (CLAP) changes the inputs in a
different way: it computes its contrastive terms on frozen text features of a
prompt, an augmented copy of it, and the class name, and it reads no images
during training [arxiv:2311.16445 §4.2].

## Contradictions

- **RSC as best.** The text says RSC is best on both benchmark datasets. The
  table shows RSC and CORAL tied on PACS at 82.0, so RSC is best only on
  Office-Home [arxiv:2103.03097 §VII.C].
- **DANN and ANDMask versus ERM.** The text says they perform worse than ERM.
  The table shows DANN above ERM on PACS (81.4 vs 81.2), and both below ERM
  on Office-Home [arxiv:2103.03097 §VII.C].
- **Meta-learning notation.** The meta update uses α as the outer step, but
  the text names η and β as the outer and inner learning rates. The survey
  points to the MAML and MLDG papers for the exact rule [arxiv:2103.03097 §IV.C.2].
- **Mapillary claim (RobustNet).** The text says ISW is best on BDD-100K,
  Mapillary and SYNTHIA when trained on Cityscapes [arxiv:2103.15597 §5.2.1].
  The Cityscapes table gives ISW 58.64 on Mapillary, below IW (58.90) and
  IRW (59.20) [arxiv:2103.15597 tab. 2].
- **IBN-Net claim (RobustNet).** The supplement says ISW beats IBN-Net for
  every number of clusters k [arxiv:2103.15597 app. A.2]. On SYNTHIA, IBN-Net
  scores 27.90 [arxiv:2103.15597 tab. 1]. Only k = 3 (28.30) is above it; k = 2,
  5, 10 and 20 score 27.70, 27.52, 27.68 and 27.47 [arxiv:2103.15597 tab. 7].
- **Fish's own tables.** The main Wilds table gives Fish 74.7 on Camelyon17 and
  72.8 on CivilComments worst-group accuracy. The appendix gives 79.5 and 74.2
  for the same method and test sets, and gives ERM 73.3 and 58.0 where the main
  table has 70.3 and 56.0 [arxiv:2104.09937 tab. 3] [arxiv:2104.09937 app. C.2]
  [arxiv:2104.09937 app. C.4]. Both places put Fish above ERM on these two
  datasets, but the size of the gap differs.
- **Gain size across papers.** The survey finds its best image method 0.8
  points above ERM [arxiv:2103.03097 §VII.C]. Fish is 0.1 points above ERM on
  DomainBed [arxiv:2104.09937 tab. 4] and several points above it on some Wilds
  datasets [arxiv:2104.09937 tab. 3]. EERM reports gains of up to
  7.2% on Facebook-100 and 8.1% relative on OGB-Arxiv [arxiv:2202.02466 §5.2]
  [arxiv:2202.02466 §5.3]. RobustNet reports gains of 7.63, 10.06 and 12.15
  points over its own baseline on C, B and M [arxiv:2103.15597 tab. 5]. The
  settings and baselines differ, so the numbers do not conflict directly. The
  survey's gains should not be read as typical for graphs or segmentation.
  CLAP's average zero-shot gain over CLIP is 2.5 points on four of the same
  benchmark datasets [arxiv:2311.16445 tab. 3], but its baseline is a
  pretrained zero-shot model, not an ERM model trained on the sources.

## Pitfalls

- **Benchmark numbers are few runs.** The survey reports one number per cell,
  with no seed count or deviation [arxiv:2103.03097 §VII.C]. RobustNet
  averages three runs but reports no spread [arxiv:2103.15597 §5.2.1]. Fish
  uses three seeds, and ten on Camelyon17 [arxiv:2104.09937 §4.2]. Treat
  small differences as noise.
- **Source accuracy can drop.** ISW lowers the source column: 72.10 against
  73.45 on GTAV [arxiv:2103.15597 tab. 1], and 76.41 against 77.51 on
  Cityscapes [arxiv:2103.15597 tab. 2]. The paper says ISW is best on every
  target set except the source [arxiv:2103.15597 §5.2.1].
- **Training cost is not measured.** The paper reports unchanged inference
  time and FLOPs [arxiv:2103.15597 tab. 6]. It does not measure training time,
  and training adds a warm-up pass, a variance estimate over photometric pairs,
  and the loss on each of three layers [arxiv:2103.15597 §5.2.3].
- **Mask depends on the statistics images.** The selected entries depend on
  which images and photometric pairs are used. The paper does not give the
  number of pairs or trials. See [the card's Paper vs code section](/papers/2103.15597).
- **Model selection uses source data only.** Under the training-domain
  protocol, the validation split comes from the sources. The survey says this
  protocol cannot always reach the best performance, while test-domain
  validation uses target data and so does not fit DG [arxiv:2103.03097 §VII.B].
- **Theory assumptions are narrow.** The DG bounds assume covariate shift and,
  for some results, a kernel or VC-dimension setting. One bound assumes the
  target lies in the convex hull of the sources [arxiv:2103.03097 §III.B].
- **Image tasks only.** The survey's benchmark covers image classification.
  RobustNet covers urban-scene segmentation. The survey lists other settings,
  but does not benchmark them [arxiv:2103.03097 §VII.A] [arxiv:2103.03097 §VII.C].
- **Dense editor memory (graph variant).** The editor's edit matrix is over
  node pairs, and the paper gives no sparse or subsampled alternative
  [arxiv:2202.02466 §3.2] [arxiv:2202.02466 app. A]. The reference code
  keeps one n × n matrix per editor. At OGB-Arxiv size (169,343 nodes
  [arxiv:2202.02466 tab. 1]) the card estimates about 115 GB per editor in
  float32. Any large-graph use
  needs a design the paper does not describe. See
  [the card's Paper vs code section](/papers/2202.02466).
- **No code reuse for the graph variant.** The authors' repository has no
  license file, so do not copy its code. Implement from the paper's equations
  [arxiv:2202.02466 §3.2].
- **Unstated graph settings.** The number of environments K, the number of
  epochs and the optimisers are not stated in the paper, and the card lists
  where the reference code's values differ from the paper's
  [arxiv:2202.02466 app. F.2] [arxiv:2202.02466 app. F.3].
- **Reference code differs from the paper.** The RobustNet repository uses
  other covariance normalisation, a different warm-up count, and per-setting
  cluster counts. Keep the code's conventions when reproducing its reported
  numbers [see the card's Paper vs code section](/papers/2103.15597).
- **Fish's update is not the paper's literal rule in the code.** Algorithm 1
  applies the outer rate directly and uses plain gradient steps in the inner
  loop [arxiv:2104.09937 alg. 1]. The released code divides the outer rate by
  the number of meta-steps and carries the ERM optimiser state across inner
  loops [see the card's Paper vs code section](/papers/2104.09937). Match the
  code if you want its reported numbers, and say which one you used.
- **The alignment holds only in a limit.** Fish drops the second-order term on
  purpose. The paper's alignment result holds as the inner rate goes to zero,
  and the paper gives no error bound at the rates it uses
  [arxiv:2104.09937 §3.4] [arxiv:2104.09937 app. A] [arxiv:2104.09937 tab. 14].
- **Fish rates differ by dataset.** Across the Wilds datasets the inner rate
  runs from 2e-6 to 1e-3 and the outer rate from 0.01 to 0.1
  [arxiv:2104.09937 tab. 14]. Copy them per dataset, not as one global value.
- **Fish speed claim is unmeasured.** The paper says Fish is about ten times
  faster than direct IDGM, but gives no timing table [arxiv:2104.09937 §4.1].
  Each meta-step costs one gradient step per source domain it visits
  [arxiv:2104.09937 alg. 1].
- **Pretraining matters for Fish too.** Several Wilds runs start from an ERM
  checkpoint, and results change with how far that checkpoint converged
  [arxiv:2104.09937 app. E] [arxiv:2104.09937 tab. 15]. Give the baseline the
  same starting point.
- **CLAP's α was chosen on evaluation data.** The paper does not say how the
  inference weight α was picked. The reference code's search scores zero-shot
  accuracy on the test environments, so the reported gains may be optimistic.
  Use a held-out split [see the card's Paper vs code section](/papers/2311.16445)
  [arxiv:2311.16445 app. A.1].
- **CLAP's OfficeHome and DomainNet settings are appendix-only.** The
  temperature of 0.3 and the class-term weight of 0.0001 for those datasets
  are in the appendix, and the shipped training config covers VLCS only
  [arxiv:2311.16445 app. A.1]. Carry the dataset values explicitly.
- **CLAP's loss differs from its equation in code.** The code uses a
  two-directional InfoNCE with same-view negatives, not the one-directional
  form in [arxiv:2311.16445 eq. 2]. With few classes, a batch repeats class
  names, and the code masks only the diagonal, so same-class pairs count as
  negatives [see the card's Paper vs code section](/papers/2311.16445).
- **No code reuse for CLAP.** The repository has no LICENSE file, so
  implement from the paper's equations [arxiv:2311.16445 §4.2].

## Papers

- [Generalizing to Unseen Domains: A Survey on Domain Generalization](/papers/2103.03097): survey of DG methods in three groups, the theory of the problem, the dataset list, and a benchmark of seven methods on PACS and Office-Home with ResNet-18.
- [RobustNet: Improving Domain Generalization in Urban-Scene Segmentation via Instance Selective Whitening](/papers/2103.15597): one-source DG for urban-scene segmentation, penalising the style-sensitive covariance entries of instance-normalised layers (ISW) during training only.
- [Gradient Matching for Domain Generalization](/papers/2104.09937): Fish, a first-order approximation of inter-domain gradient matching that runs one inner gradient step per domain on a copy of the weights; evaluated on Wilds and DomainBed.
- [Handling Distribution Shifts on Graphs: An Invariance Perspective](/papers/2202.02466): node-level DG on graphs, with learned edge edits that create virtual environments (EERM) and a mean-plus-variance risk objective.
- [CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts](/papers/2311.16445): a text-only residual MLP on frozen CLIP features, trained contrastively on prompts and style-dropped copies, then reused on image features at inference.