learn

Face Swap Glossary

Face swap glossary: deepfake, persona, virtual camera, inference latency, alignment, and 20+ live face swap terms defined for creators and developers.

Part of our learn hub.

Core terminology mapConcept diagram explaining Live face swap vocabulary.Core terminology mapLive face swap vocabularyPersonaSwap identityON AIRLive sessionVCamVirtual cameraCreditsLive minutes
Core terminology map
Core terminology mapConcept diagram explaining Live face swap vocabulary.

Clear vocabulary prevents expensive mistakes, buying GPU hardware for cloud swap, expecting filters to replace personas, or confusing offline deepfakes with live virtual camera output. This glossary defines terms used across LiveSwap docs, streaming forums, and research papers. Part of our education articles.

A–Z terms

Alignment

Face alignment maps detected landmarks (eyes, nose, mouth corners) to a canonical coordinate frame so the swap model knows how to warp texture onto your current pose. Poor alignment causes floating jaws and eye drift. Live webcam angle must reasonably match persona photo framing, see source photo guide.

AR filter

Augmented reality face filter, in-app effect (Snapchat lens, Instagram filter) attaching meshes or textures to tracked face mesh inside one platform. Distinct from photorealistic cross-app live swap, swap vs filter guide.

Avatar

Generic term for on-screen character representation. May mean VTuber rig, game skin, or live swap persona. LiveSwap UI uses "avatar" for uploaded photo slots; docs prefer persona for identity consistency, /personas.

Batch / offline swap

Processing recorded video or images non-real-time. Output is file, not live camera. High quality, high latency, opposite of live face swap.

Browser source (OBS)

OBS input type rendering a web page, used to display LiveSwap stage in OBS before Virtual Camera export. Setup: webcam setup guide.

Cloud inference

Running neural networks on remote GPU servers instead of local hardware. LiveSwap model, no CUDA on your machine. Trade network for convenience, face swap deployment guide.

Credits (LiveSwap)

Billing unit: 1 credit = 1 live minute of ON AIR swap, metered to the second. Uploads and prep free. Plans at /pricing.

Deepfake

Colloquial term for AI-altered faces in media, often offline video, often associated with impersonation harm. Technology overlaps live swap; ethics and pipeline differ, history of face swap.

Detection

Automated or human identification of manipulated faces via temporal artifacts, sync errors, classifiers. Live swap detectability varies, can face swaps be detected.

End-to-end latency

Total delay from real movement to swapped pixel display including capture, network, inference, OBS, meeting app. Target sub-500ms, latency breakdown.

Face detection

Locating face bounding box and landmarks in each frame. Failure modes: bad lighting, occlusion, sunglasses, swap failure guide.

Faceless creator

Content creator omitting real identity on camera, voice-only, avatar, or live swap persona. Trend context: rise of faceless creators.

GAN (Generative Adversarial Network)

ML architecture where generator and discriminator compete, historically improved synthetic face realism. Modern swap pipelines may use GANs plus other architectures; consumers need not specify model family to use LiveSwap.

GPU inference

Running swap models on local graphics card, DeepFaceLive path. Requires NVIDIA CUDA in typical community setups, hardware checklist.

Inference

Forward pass of trained model producing swapped output pixels from aligned input. Measured in milliseconds per frame on capable hardware.

Jitter

Network delay variance, causes stutter distinct from constant lag. Ethernet reduces jitter vs Wi-Fi, delay troubleshooting.

Landmark

Semantic points on face (eye corners, lip top/bottom) used for alignment and expression transfer.

Latency

See end-to-end latency. Critical perceptual metric for live conversation.

Live face swap / real-time face swap

Continuous swap on webcam or capture feed outputting to virtual camera or stream while performer acts, definition: what is swap guide.

Model / checkpoint

Trained neural network weights file. Desktop users swap ONNX checkpoints manually; LiveSwap serves models server-side.

NCII (non-consensual intimate imagery)

Illegal harmful deepfake category, distinct from consented persona streaming. Prohibited on LiveSwap, usage terms.

ONNX

Open Neural Network Exchange format, common for portable models in DeepFaceLive ecosystem.

ON AIR (LiveSwap)

Session state when live swap output active and credits meter. Distinct from dashboard prep.

Persona

Consistent target identity from uploaded source photos mapped onto performer landmarks. Encrypted storage; delete anytime. Photo guide: avatar requirements.

Photorealistic swap

Swap aiming to match natural skin texture and lighting, vs stylized cartoon filter.

Pipeline

Full sequence capture → detect → align → infer → encode → output. Explained: inference overview.

Post-production face swap

Swap on recorded files, Deepswap-style, not live virtual camera category LiveSwap occupies.

Resolution tier

Output pixel dimensions capped by plan, Basic 480p, Creator 720p, Pro/Studio 1080p.

Responsible use

Ethical framework: consent, no impersonation fraud, platform compliance, ethical use guide.

RTMP / streaming protocol

How OBS sends video to Twitch/YouTube, downstream of swap output; swap quality must survive encoder settings on OBS live swap.

Source photo / target identity

Reference image teaching model appearance, must be consented likeness.

Sub-500ms

LiveSwap latency target under good network, not guaranteed on all connections.

Temporal coherence

Frame-to-frame stability, avoiding flicker at cheeks and hairline. Harder live than offline batch.

Uncanny valley

Perceptual discomfort when synthetic face almost but not quite human, mitigated via lighting, photo, lag fixes, natural face swap tips.

Upload bandwidth

Outbound Mbps limiting cloud swap, primary "hardware" constraint for LiveSwap.

Virtual camera

Software device appearing in Zoom/OBS camera lists, carries swapped feed. OBS Start Virtual Camera common path.

VTuber

Creator using animated 2D/3D avatar driven by tracking, alternative to photorealistic neural swap; sometimes combined in hybrid shows.

WebRTC

Browser real-time communication stack for camera capture and low-latency media, LiveSwap browser transport layer.

x264 / NVENC

Software vs hardware video encoders in OBS, overload adds latency to swap chain.

Related guides → /learn hub

Explore clusters from learn index:

Start building: start streaming guide. Compare tools: comparison roundup.

Definitions align with PRODUCT-FACTS, do not infer unlisted features from glossary alone.

Additional terms (extended)

Agora / WebRTC session, Real-time transport layer carrying camera frames between browser and cloud inference in LiveSwap architecture. Distinct from RTMP streaming to Twitch.

Batch inference, Processing many frames or files with seconds to minutes per unit; offline deepfake and Deepswap-style tools. Opposite of live per-frame budget.

Blend mask, Region defining where swapped face merges with performer neck and hair; quality issues often show at mask boundary.

BRB screen, Stream offline graphic; stop LiveSwap ON AIR during BRB to avoid burning credits on static slides.

Capture card, Hardware ingesting HDMI from gaming PC to streaming PC; swap runs on streaming PC in dual-PC setups.

Consent layer, Ethical requirement separate from technology; persona source, participants, and audience expectations.

Credit, LiveSwap: one live minute ON AIR. Other vendors: may mean file render units, always read pricing footnotes.

DeepFaceLab, Offline training ecosystem feeding some local live tools; not interchangeable with LiveSwap persona upload.

Encoder lag, OBS dropped frames from x264 or NVENC overload; mimics swap lag but fixed via encoder settings.

Face mesh, AR filter landmark structure; lighter than full neural identity swap.

Green room, Pre-show webinar space; test persona before public go-live.

Identity replacement, Swapping whose face appears while preserving performer motion; core live swap goal.

Keyframe interval, Twitch requires 2-second GOP; OBS output setting unrelated to swap but mandatory for ingest.

Landmark detection, Finding eyes, nose, mouth each frame; fails in profile or heavy shadow.

Metering, LiveSwap bills ON AIR seconds; prep and uploads excluded.

NDI, Network video transport between PCs; optional dual-PC streaming path.

Non-consensual intimate imagery (NCII), Primary legal enforcement target for malicious deepfakes; distinct from consented persona streaming.

ON AIR, LiveSwap session state when swap output active and credits consume.

Paging file, Windows virtual memory DeepFaceLive users configure; irrelevant to cloud browser swap.

PiP, Picture-in-picture face cam overlay on gameplay streams.

QoS, Router quality-of-service; may prioritize upload for swap stability.

Responsible disclosure, Telling audiences or co-hosts when synthetic appearance matters to trust.

RTMP, Protocol OBS uses to send composed video to Twitch or YouTube Live.

Second-precision billing, LiveSwap credits round to live seconds, not rounded-up minutes.

Synthetic media, Umbrella term platforms use for AI-altered video; includes swap, filters, and generated avatars.

Threat model, Determines cloud vs local choice for journalists and security-sensitive users.

Touch up my appearance, Zoom beauty filter; disable when using swap to avoid double processing.

Upload headroom, Sustainable upstream Mbps for chosen resolution tier.

VTuber rig, Animated avatar alternative to photoreal swap; different aesthetic and setup cost.

Wire fraud, Criminal context when swap impersonates executives to steal money; zero tolerance under /legal/aup.

Using this glossary in team wikis

Copy terms into internal documentation with links back to Learn hub articles for depth. Update quarterly as platform policies and product tiers evolve, verify see plan options before citing minute caps in contracts.

Acceptable use policy (AUP)

LiveSwap rules at /legal/aup, consented personas, no impersonation, no NCII, no fraud. Distinct from criminal law but binding on product use.

Autoencoder (face swap context)

Neural architecture compressing identity into latent code and reconstructing face, historical training approach for offline swaps; conceptual ancestor of modern live models.

Bitrate

Data rate of encoded video stream, higher bitrate reduces compression artifacts but increases upload time for cloud swap paths.

Blending mask

Soft alpha region defining where synthesized face merges into original frame, poor masks cause visible seams.

Capture card

Hardware ingesting HDMI camera feed into PC, used in dual-PC streams; adds small latency vs direct USB webcam.

Consent

Permission to use likeness in persona photos and to present synthetic identity to audiences, responsible face swap.

DeepFaceLive (DFL)

Open-source Windows-centric real-time local swap ecosystem, DeepFaceLive alternative.

Expression coefficients

Numerical parameters describing face pose and expression, swap models map source coefficients onto target appearance.

Frame dropping

Skipping frames when inference slower than capture, causes stutter distinct from constant latency.

Generative model

Neural network synthesizing pixels, includes GANs, diffusion variants in research; LiveSwap uses production-optimized inference graphs server-side.

HDMI loop

Wiring pattern feeding camera through capture device, adds hardware latency budget line item.

Identity embedding

Vector representation of target face learned from persona photos, core of consistent character across sessions.

Impersonation

Presenting as real identifiable person without permission, prohibited /legal/aup.

Live minutes

Metered ON AIR duration on LiveSwap, 1 credit = 1 minute.

MediaStream API

Browser API accessing camera/microphone, LiveSwap capture entry point.

Morphable model

3D statistical face model, academic precursor influencing alignment thinking.

NVENC

NVIDIA hardware encoder in OBS, reduces CPU encode latency vs x264 when configured correctly.

Occlusion

Objects covering face region (hand, mic, hair), breaks landmark detection.

Poisson blending

Image processing technique merging seams, some swap pipelines use similar feathering concepts.

PoP (point of presence)

Geographic datacenter location for cloud inference, distance affects RTT latency.

rPPG

Remote photoplethysmography, pulse detection from skin color; research detection signal, weak on compressed live video.

RTT (round-trip time)

Network ping time there-and-back, doubles impact on cloud swap vs one-way upload estimates.

Session prep

Upload persona, test camera, configure OBS before ON AIR, does not consume live credits on LiveSwap.

Synthetic media

Regulatory/policy term for AI-altered audio/video, overlaps deepfake colloquial usage.

Two-party consent

Recording law requiring all parties agree in some US states, applies to swapped calls too.

Virtual camera driver

OS-level software exposing processed video as webcam, OBS Virtual Camera, vendor bridges.

WebGPU

Browser GPU API, future local inference experiments; LiveSwap uses cloud inference today.

Worked glossary exercise

Translate: "DFL mask slip at 720p with high RTT" → DeepFaceLive alignment failure at HD resolution exacerbated by network round-trip delay; consider lag troubleshooting or cloud vs local swap migration.

Using this glossary in documentation

Policy teams, excerpt AUP, consent, impersonation, NCII entries.

Engineers, virtual camera, pipeline, latency, WebRTC cross-links to guides.

Creators, persona, photorealistic swap, faceless creator before purchasing plans.

Return to knowledge base for narrative depth beyond definitions.

Autoencoder

Neural network that compresses input into a latent representation and reconstructs output. Classic face swap training paired encoders/decoders for identity A and B, swap by decoding with wrong identity code. Modern live products hide this complexity; understanding autoencoders explains why source photo quality anchors identity.

Bitrate

Data rate of compressed video, measured in Mbps or Kbps. Higher resolution swap tiers need higher bitrate and upload bandwidth. OBS encoder settings interact with swap output sharpness.

Blending / compositing

Merging synthesized face pixels onto original frame using soft masks and color matching. Poor blending creates visible halos at jaw and hairline, often lighting issue, not model issue. realism tuning guide.

Browser-based face swap

Capture and session control in web browser; inference may run locally (WebGPU experiments) or in cloud (LiveSwap). LiveSwap: no install, no CUDA.

Clap test

Informal sync diagnostic, clap hands on camera, compare audio spike to visible contact in recording. Detects lip-sync lag from network or OBS buffer misconfiguration.

Consent

Permission to use a likeness or present synthetic identity to an audience. Layers: photo subject, call participants, viewers expecting authenticity. responsible-use policy.

CUDA

NVIDIA parallel computing platform. Required for typical local GPU swap stacks (DeepFaceLive). Not required for LiveSwap cloud path.

DeepFaceLive (DFL)

Open-source desktop real-time face swap using local NVIDIA GPU. High customization; steep setup. local GPU swap alternative.

Expression transfer

Mapping performer's facial motion (mouth, brows, eyes) onto persona appearance. Core mechanism of live swap, you act, viewer sees persona act.

Frame buffer

Temporary frame storage in OBS or video apps. Larger buffers smooth stutter but add latency, bad for conversational lip sync.

Identity consistency

Stable persona appearance across sessions and platforms. LiveSwap persona library goal, audience recognizes host face over months.

Impersonation

Presenting as a specific real identifiable person without permission. Prohibited on LiveSwap per /legal/aup. Distinct from original character persona.

Lip sync

Alignment of mouth movement with audio. Desync makes photoreal swap feel fake, fix latency before resolution. reduce swap delay.

Live minutes

Metered duration when LiveSwap swap pipeline is ON AIR, billed via credits. Distinct from login time or free upload time.

Mask slip

Artifact where swapped texture drifts off face during head motion, alignment failure, latency spike, or lighting change.

MediaStream API

Browser API accessing webcam and microphone. Foundation of LiveSwap browser capture.

NDI (Network Device Interface)

Video-over-IP protocol; some advanced streamers use NDI instead of virtual camera. LiveSwap standard path: OBS Browser Source + Virtual Camera.

ONNX Runtime

Engine running ONNX model files, common in desktop swap tooling. Cloud users rarely interact directly.

Persona library

Encrypted storage of uploaded persona photos in LiveSwap. Users delete anytime. Marketing: /personas.

Provenance (C2PA)

Metadata attesting how media was created or edited, regulatory vocabulary for synthetic content; more relevant to uploaded clips than ephemeral live.

Reenactment (face reenactment)

Research area: one person's expressions drive another's appearance. Live swap is consumer reenactment, performer drives persona.

Segmentation mask

Map defining which pixels are face, hair, background. Mask errors at hairline cause visible seams.

Synthetic media

Regulatory and platform term for AI-altered or AI-generated audiovisual content, overlaps "deepfake" but used in policy documents.

Training vs inference

Training learns model weights offline over hours/days. Inference applies trained model to live frames in milliseconds. LiveSwap users infer; provider manages training/adaptation to persona photos.

Virtual background

Platform feature replacing room behind user. Can fight swap edge blending, often disable when using swapped output in Zoom/Meet.

WebGPU

Browser API for local GPU compute. Research path for client-side inference; LiveSwap uses cloud inference instead.

Zero-shot swap

Swap toward target identity without per-person hours of training, modern products infer from few uploaded photos. LiveSwap persona upload pattern.

Common confusions (quick reference)

Term ATerm BDifference
PersonaFilter lensPersona persists across apps; filter is in-app ephemeral
InferenceTrainingInference applies model live; training builds model offline
Virtual cameraScreen captureVirtual cam is a device; screen capture films a window
Live swapFile deepfakeLive is continuous webcam; file is post-production
Cloud swapLocal swapWhere GPU runs, network vs CUDA
DeepfakeFace swapDeepfake is broad harmful-association term; swap is specific technique
VTuberLive swapStylized rig vs photoreal neural swap
ON AIRUpload prepON AIR meters credits; prep is free

Glossary usage for teams

Policy leads excerpt Acceptable use, Impersonation, and Consent entries for internal wiki.

Engineers cross-link Virtual camera, Pipeline, Latency, and WebRTC when documenting OBS integration.

Creators read Faceless creator, Persona, and Photorealistic swap before purchasing plans.

Ethics precede terminology, read face swap ethics and /legal/aup before going live.

Frequently asked questions

Start your first live face swap

No install, no GPU. Upload a photo, pick a persona, and go live in minutes.