AI face swap glossary

Also called: AI face swap, face swapping

A computer-vision task that transfers identity cues from a source face into a target image or video while attempting to preserve target attributes such as pose, expression, hair, and background.

Evidence boundary: This is a task definition, not a guarantee that every system preserves every attribute or that a particular use has permission.

Primary sources: PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion; Fine-Grained Face Swapping via Regional GAN Inversion

Use it in context: Photo face swap guide; Verification methodology

Also called: source identity, identity reference

The person or reference image that supplies the identity cues intended for the generated face.

Evidence boundary: The source does not normally define the target scene, body, hair, pose, or background, and the exact split depends on the method.

Primary sources: PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion; Reinforced Disentanglement for Face Swapping without Skip Connection

Use it in context: Photo input guide; Input readiness checker

Also called: target image, target video

The image or video that supplies the scene and the non-identity attributes a face-swap method attempts to retain.

Evidence boundary: Methods differ in which target attributes they retain, so target preservation must be evaluated rather than assumed.

Primary sources: PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion; DynamicFace: High-Quality and Consistent Face Swapping for Image and Video

Use it in context: Video face swap guide; Quality scorecard

Also called: face localisation, face localization

The process of locating one or more face regions in an image or video frame before alignment, feature extraction, or generation.

Evidence boundary: A detected face region is not an identity decision, a consent check, or proof that later processing will succeed.

Primary sources: RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild

Use it in context: Multiple face swap guide; Pipeline architecture

Also called: face landmarks, landmark points

Estimated facial key points, such as eye corners, nose, and mouth locations, used to describe facial geometry for alignment or conditioning.

Evidence boundary: Landmark layouts, point counts, and error units vary by detector and protocol; they are not directly interchangeable.

Primary sources: RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild; DiffSwap: High-Fidelity and Controllable Face Swapping via 3D-Aware Masked Diffusion

Use it in context: Input readiness checker; Technical glossary

Also called: facial alignment, aligned face crop

A geometric normalization step that uses detected facial structure to place a face into a consistent crop, scale, and orientation for later processing.

Evidence boundary: Alignment conventions and transforms vary, and an aligned crop does not establish identity, quality, or permission.

Primary sources: RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild; ArcFace: Additive Angular Margin Loss for Deep Face Recognition

Use it in context: Pipeline architecture; Photo face swap guide

Also called: identity embedding, facial feature vector

A learned numerical representation designed so that facial images can be compared in a feature space under a specific recognition model and preprocessing protocol.

Evidence boundary: Vector dimensions, distance functions, and decision thresholds depend on the model and evaluation setup; an embedding is not inherently an identity verdict or mathematically guaranteed to be non-reconstructable.

Primary sources: ArcFace: Additive Angular Margin Loss for Deep Face Recognition

Use it in context: Quality scorecard; Privacy and consent guide

Also called: identity retention, identity preservation score

A protocol-dependent comparison between source-identity and output representations, commonly computed from a named face encoder and similarity function.

Evidence boundary: Scores are not portable across encoders, crops, datasets, thresholds, or demographics and should not be presented as a biometric verdict for one output.

Primary sources: ArcFace: Additive Angular Margin Loss for Deep Face Recognition; PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion

Use it in context: Face swap quality metrics; Research catalog

Also called: attribute preservation, target consistency

The extent to which a result retains intended target characteristics such as pose, expression, hair, lighting, clothing, and background while changing facial identity cues.

Evidence boundary: The evaluated attributes and estimators must be named; one aggregate score can conceal failures in an individual attribute.

Primary sources: PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion; DynamicFace: High-Quality and Consistent Face Swapping for Image and Video

Use it in context: Output quality scorecard; Photo quality guide

Also called: swap mask, facial mask

A spatial map that identifies which image region can be generated, blended, or evaluated as part of a face-swap operation.

Evidence boundary: Mask construction and semantics vary by method; a mask does not by itself solve occlusion, boundary, or color inconsistency.

Primary sources: DiffSwap: High-Fidelity and Controllable Face Swapping via 3D-Aware Masked Diffusion; Fine-Grained Face Swapping via Regional GAN Inversion

Use it in context: Pipeline architecture; Output review guide

Also called: facial semantic segmentation

Pixel-level semantic segmentation of facial components and surrounding regions, such as skin, eyes, mouth, hair, or background.

Evidence boundary: Label sets and segmentation quality depend on the parser and dataset, and incorrect regions can propagate into later synthesis or compositing.

Primary sources: Fine-Grained Face Swapping via Regional GAN Inversion

Use it in context: Photo quality guide; Pipeline architecture

Also called: boundary blending, face compositing

The compositing step that transitions a synthesized facial region into the target around skin tone, edges, lighting, and nearby content.

Evidence boundary: A visually smooth boundary does not establish correct identity, natural internal features, or consistent behavior across video frames.

Primary sources: Fine-Grained Face Swapping via Regional GAN Inversion; Reinforced Disentanglement for Face Swapping without Skip Connection

Use it in context: Output quality scorecard; Photo face swap guide

Also called: face obstruction, occluded face

A condition in which part of a face is hidden by hair, hands, glasses, objects, another person, or the image boundary.

Evidence boundary: Occlusion handling is method- and scene-dependent; a benchmark must specify the affected region and viewing conditions.

Primary sources: Fine-Grained Face Swapping via Regional GAN Inversion; DynamicFace: High-Quality and Consistent Face Swapping for Image and Video

Use it in context: Input readiness checker; Video preparation guide

Also called: face pose, head orientation

The estimated orientation of a head relative to the camera, often represented by yaw, pitch, and roll or by landmarks and 3D parameters.

Evidence boundary: Pose values depend on the estimator and coordinate convention; matching pose alone cannot predict output realism.

Primary sources: DiffSwap: High-Fidelity and Controllable Face Swapping via 3D-Aware Masked Diffusion; PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion

Use it in context: Photo quality guide; Input readiness checker

Also called: expression preservation

The visible configuration of facial features associated with movements such as smiling, blinking, or opening the mouth, often treated as a target attribute in face swapping.

Evidence boundary: Expression labels and parameters are estimator-dependent and can miss subtle, asymmetric, or culturally interpreted behavior.

Primary sources: PixSwap: High-Resolution Face Swapping for Effective Reflection of Identity via Diffusion; DynamicFace: High-Quality and Consistent Face Swapping for Image and Video

Use it in context: Output quality scorecard; Video face swap guide

Also called: temporal coherence, frame-to-frame consistency

The stability of identity cues, target attributes, and visual detail across consecutive video frames rather than in one isolated frame.

Evidence boundary: A high consistency measure can preserve the wrong result consistently, so identity correctness and visible quality must be evaluated separately.

Primary sources: DynamicFace: High-Quality and Consistent Face Swapping for Image and Video; CanonSwap: High-Fidelity and Consistent Video Face Swapping via Canonical Space Modulation

Use it in context: Video face swap guide; Video quality scorecard

Also called: motion field, pixel correspondence

An estimated dense motion field that maps image locations between frames and can support motion analysis, warping, or temporal diagnostics.

Evidence boundary: Optical flow is an estimate that can fail around occlusion, blur, lighting changes, and newly visible regions; it is not a quality score by itself.

Primary sources: RAFT: Recurrent All-Pairs Field Transforms for Optical Flow; CanonSwap: High-Fidelity and Consistent Video Face Swapping via Canonical Space Modulation

Use it in context: Video workflow guide; Pipeline architecture

Also called: SSIM

A full-reference image similarity measure that compares luminance, contrast, and structure between a test image and a reference image.

Evidence boundary: SSIM requires a meaningful aligned reference, is not face-specific, and its value depends on scale, windowing, color handling, and preprocessing.

Primary sources: Image Quality Assessment: From Error Visibility to Structural Similarity

Use it in context: Face swap quality metrics; Research methodology

Also called: LPIPS, learned perceptual similarity

A perceptual distance based on deep network features that was evaluated against human judgments of image-patch similarity.

Evidence boundary: LPIPS is not an identity score; values depend on the selected network, calibration, image preparation, and comparison protocol.

Primary sources: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

Use it in context: Face swap quality metrics; Controlled research

Also called: FID

A set-level distance between feature distributions from generated and reference image collections, originally computed with Inception features.

Evidence boundary: FID is not a defensible single-image score, and results depend on sample count, feature extractor, preprocessing, reference set, and implementation.

Primary sources: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium; Rethinking FID: Towards a Better Evaluation Metric for Image Generation

Use it in context: Face swap quality metrics; Research catalog

Also called: NR-IQA, blind image quality assessment

A method that estimates visual quality from an output without requiring a paired pristine reference image.

Evidence boundary: Its validity depends on training data and validation design; a score can fail on unseen methods, content, or distortions.

Primary sources: Rank-based No-reference Quality Assessment for Face Swapping

Use it in context: Output quality scorecard; Verification methodology

Also called: human evaluation rubric, output review rubric

A repeatable rubric that asks a reviewer to record visible criteria, rating anchors, critical gates, and notes for a specific output.

Evidence boundary: It records structured human observations, not model accuracy, biometric identity, a probability, or a provider-wide ranking.

Primary sources: DeepSwapAI face swap output quality scorecard; Rank-based No-reference Quality Assessment for Face Swapping

Use it in context: Open the scorecard; Download research assets

Also called: Coalition for Content Provenance and Authenticity

An open technical standard for recording tamper-evident provenance information about digital content through signed manifests, assertions, ingredients, and claims.

Evidence boundary: C2PA can provide provenance evidence and tamper signals, but the specification does not determine whether depicted content is true, accurate, or ethically permitted.

Primary sources: C2PA Technical Specification 2.4; C2PA Explainer 2.3

Use it in context: Trust and safety; Verification methodology

Also called: C2PA Content Credentials, content provenance credentials

The user-facing presentation of provenance information associated with a digital asset through the C2PA ecosystem.

Evidence boundary: Availability and displayed details depend on the credential, signer, validator, asset, and platform; a credential is not a declaration that content is factually true.

Primary sources: C2PA Explainer 2.3; C2PA Technical Specification 2.4

Use it in context: Trust center; Privacy and consent guide