When we train a generative model to perform style transfer or domain adaptation—such as translating a photo into a cartoon, changing its historical era, or adapting a surveillance image to a new camera network—the model's primary objective is distribution matching. Generative Adversarial Networks (GANs) and diffusion models excel at capturing global statistics like color palettes, textures, and background illumination.
However, this optimization creates a fundamental conflict. Identity is defined by highly specific, fine-grained, and low-entropy structural features: the precise geometry of a face, the exact spacing of facial features, or the unique patterns on a person's clothing. Because standard generative objectives prioritize global visual realism and style matching, they often discard or distort these fine-grained details to satisfy the domain discriminator. As noted by Weijian Deng et al. in their work on Image-Image Domain Adaptation With Preserved Self-Similarity and Domain-Dissimilarity for Person Re-Identification, without explicit identity constraints, the translation process corrupts the underlying discriminative features, making the generated data useless for downstream recognition tasks. This challenge spans multiple domains, from cross-dataset person re-identification (Cross-Dataset Person Re-identification Using Similarity Preserved Generative Adversarial Networks) to artistic portrait generation (RoboCoDraw: Robotic Avatar Drawing with GAN-Based Style Transfer and Time-Efficient Path Optimization; CSGO: Content-Style Composition in Text-to-Image Generation).
To resolve this tension, researchers have developed three core mechanisms to protect identity while allowing style to fluidly change: latent anchoring, spatial isolation, and disentangled feature injection.
Latent Anchoring via Identity Mapping Losses
To prevent the generator from drifting into generic representations of the target domain, researchers introduce "latent anchors." The core intuition is that if we pass both the original image and the stylized image through a deep feature extractor, their representations in an identity embedding space should remain close.
This is operationalized through an identity mapping loss or content-preserving constraints. For instance, in Semantic Consistency and Identity Mapping Multi-Component Generative Adversarial Network for Person Re-Identification, the authors introduce an identity mapping loss specifically to ensure that an individual's discriminative features survive cross-domain translation.
Similarly, the authors of Non-Visual to Visual Translation for Cross-Domain Face Recognition integrate a content-preserving Siamese network into their style-transfer pipeline. By forcing the network to minimize the distance between the source and translated features in a shared embedding space, the model learns that while textures and lighting must change, the underlying structural identity must remain invariant. Zhiqin Zhu et al. leverage a similar Siamese architecture in A GAN-Based Self-Training Framework for Unsupervised Domain Adaptive Person Re-Identification to bridge the domain gap without destroying pedestrian identity.
Spatial Isolation and Foreground Masking
While latent losses constrain the network in embedding space, generators can still find adversarial ways to cheat—for example, by hallucinating style elements onto the subject's face or body to satisfy the discriminator. To enforce stricter geometric preservation, researchers turn to spatial isolation.
In tasks like person re-identification, the domain gap is primarily driven by environmental factors: camera angles, background scenery, and lighting conditions. The subject's foreground identity, however, should ideally remain constant. To exploit this, Longhui Wei et al. propose Person Transfer GAN to Bridge Domain Gap for Person Re-Identification, which utilizes segmentation-based masking to protect foreground identity features during image-to-image translation.
By isolating the subject, the network is forced to learn domain-specific background and lighting characteristics without altering the core identity labels. This concept is mirrored by Haotian Tang et al. in Unsupervised Person Re-Identification With Iterative Self-Supervised Domain Adaptation, where a masked consistency loss ensures that the generative model focuses its stylistic transformations on the environment rather than corrupting fine-grained pedestrian details (Cross-Domain Person Re-Identification Based on Feature Fusion). Maintaining this identity consistency during style-based data augmentation provides a far more reliable training signal for unsupervised domain adaptation than generic style transfer (Augmented Discriminative Clustering for Domain Adaptive Person Re-Identification).
Disentangled Style Injection and Identity Mirrors
Masking works well when the foreground identity can remain visually identical, but it fails when the identity must adopt the style itself—such as turning a real face into a 1920s historical portrait (What's in a Decade? Transforming Faces Through Time), a 3D artistic model (One-shot Generative Domain Adaptation in 3D GANs), or an oil painting via text-to-image diffusion. In these scenarios, the identity cannot simply be shielded; it must be decoupled from the style and then recombined.
To solve this, researchers use style injection mechanisms like Adaptive Instance Normalization (AdaIN) and mixed attention to create "identity mirrors." For example, Yongxing Dai et al. develop a framework in Bridging the Source-to-target Gap for Cross-domain Person Re-Identification with Intermediate Domains that maps original identities into synthetic intermediate domains. By transferring only the style statistics of the target domain onto the structural features of the source, they create identity-preserving mirrors that allow downstream models to focus purely on cross-domain variations.
In the context of modern diffusion models, Yi Wu et al. present Infinite-ID: Identity-Preserved Personalization via ID-Semantics Decoupling Paradigm, which explicitly decouples facial identity features from text semantics during training. During inference, their mixed attention mechanism aligns the global style of the target prompt with the reference facial features. This ensures that the structural geometry of the face remains completely recognizable, even as its texture, color, and lighting adapt entirely to drastic artistic styles.
Key Foundational Work
Unsupervised Person Re-Identification With Iterative Self-Supervised Domain Adaptation (2019) — contributed 2 concept elements to this direction.
Image-Image Domain Adaptation With Preserved Self-Similarity and Domain-Dissimilarity for Person Re-Identification (2018) — contributed 2 concept elements to this direction.
Cross-Dataset Person Re-identification Using Similarity Preserved Generative Adversarial Networks (2018) — contributed 1 concept elements to this direction.