The Anatomy of a Convincing Deepfake Video

Not all deepfakes are created equal. The crude face-swaps that circulated in the early years of generative adversarial networks — characterised by flickering boundaries, warped geometry, and obvious blurring — have given way to a new generation of synthetic video that is perceptually indistinguishable from authentic footage at normal viewing distances and resolutions. Understanding what makes a deepfake convincing is essential for both detection and defence: the same properties that fool human perception are the properties that forensic analysis must target. This article dissects the technical anatomy of a convincing deepfake across seven dimensions, explaining both how each element contributes to perceived authenticity and how forensic analysis can identify the residual artefacts that even the most sophisticated synthesis cannot fully eliminate.

Dimension 1: Face Rendering Quality

The face is the focal point of any deepfake, and face rendering quality is the single most important factor in perceived authenticity. Modern face-synthesis models achieve remarkable visual fidelity, but several sub-components of face rendering reveal the current boundary between convincing synthesis and forensically detectable artefacts:

Resolution and Detail Hierarchy

A convincing face must reproduce detail at multiple spatial scales simultaneously: the macro-scale geometry of facial proportions, the meso-scale texture of skin surface (wrinkles, pores, fine hair), and the micro-scale structure of individual pores and skin cells. Current state-of-the-art models excel at macro-scale geometry but show progressive degradation at smaller scales.

Forensic analysis exploits this scale-dependent fidelity degradation. At macro scale, deepfake faces match their targets with sub-millimetre accuracy. At meso scale (skin texture), the synthesis typically reproduces the statistical properties of skin texture (correct spatial frequency distribution) but not the specific texture — individual pores are generated procedurally rather than reproduced from the target. At micro scale, synthesis models typically produce smooth, detail-free surfaces where authentic skin would show structure.

The practical forensic approach involves spatial frequency decomposition: separating the image into frequency bands and comparing the energy distribution in each band against the expected distribution for the apparent camera, lighting, and skin type. Deepfakes characteristically show energy deficit in the highest frequency bands (where pore-level detail should be present) and slight energy excess in mid-frequency bands (where the model's texture generation overcompensates for the loss of fine detail).

Skin Pore Rendering

Human skin pores are not randomly distributed — they follow patterns determined by anatomy, skin type, and age. Pore density varies across the face (highest on the nose and inner cheeks, lowest on the forehead and jaw), and pore orientation follows the direction of underlying muscle fibres. Current synthesis models typically generate pore-like texture noise but fail to reproduce these anatomically determined distribution and orientation patterns.

Forensic pore analysis examines both pore density maps (are the spatial statistics correct?) and pore orientation coherence (do pores follow anatomically plausible orientations?). In the most convincing deepfakes, the pore density may be roughly correct but the orientation is random — a subtle but reliably detectable artefact when examined at sufficient magnification. Our forensic modules include automated skin texture analysis that evaluates these properties.

Specular Highlights

Human skin is a multi-layered optical medium. Light striking the skin partially reflects from the surface (specular reflection) and partially penetrates, scatters through the subsurface layers, and re-emerges (subsurface scattering). The balance between specular and diffuse reflection varies across the face (oilier regions like the T-zone show stronger specular highlights) and changes with head angle relative to the light source.

Convincing deepfakes must reproduce both the specular highlight positions (which are determined by the geometry of the face relative to the light source) and the specular highlight characteristics (sharpness, intensity, spectral content). Most current models handle specular position adequately but produce highlights that are too broad (Gaussian-blurred rather than sharply defined) and that lack the correct spectral characteristics (specular reflections should be approximately the colour of the light source, not the colour of the skin).

Dimension 2: Temporal Coherence

A convincing deepfake must maintain consistency across frames — the face must look like the same face from frame to frame, skin texture should persist rather than regenerating, and the geometric relationship between facial features must remain anatomically consistent through head movements.

Temporal coherence is one of the most challenging aspects of deepfake synthesis because most generative models process frames independently or with limited temporal context. Even models that incorporate temporal attention mechanisms produce subtle frame-to-frame inconsistencies that are imperceptible at video playback speed but detectable through forensic analysis:

• Texture flickering — When individual frames are examined in sequence, the fine skin texture (pores, fine lines) does not persist consistently. Instead, the texture is regenerated per-frame, producing a subtle flickering effect visible in frame-difference analysis. Authentic video shows texture changes only due to lighting variation, not texture regeneration.
• Identity drift — Over longer sequences, the face may undergo subtle geometric changes as the model's generation process introduces small random variations. The interocular distance, nose width, or jawline shape may vary by 1–2% across a 60-second clip. These variations are below the threshold of human perception but detectable through biometric tracking.
• Asymmetry instability — Authentic faces have consistent bilateral asymmetry (one eye slightly larger, one brow slightly higher). In deepfakes, this asymmetry pattern may change frame-to-frame as the model independently generates each side of the face from its latent representation. Forensic analysis tracks asymmetry metrics over time and flags instabilities.

Dimension 3: Audio-Visual Synchronisation

The temporal alignment between lip movements and speech sounds is governed by the biomechanics of the vocal tract. Plosive consonants (/p/, /b/, /t/, /d/, /k/, /g/) require specific articulatory configurations that are visible on video. Bilabial plosives (/p/, /b/) require complete lip closure. Labiodental fricatives (/f/, /v/) require lower lip contact with upper teeth.

Convincing deepfakes achieve approximate lip-sync — the mouth opens and closes in rough temporal alignment with the speech. However, the phoneme- level precision required for complete auditory-visual coherence is extremely difficult to achieve because it requires modelling the articulatory dynamics of the vocal tract, not just the acoustic features of the speech signal.

The forensic approach involves phoneme-aligned analysis: extracting phoneme boundaries from the audio track using forced alignment and comparing the expected articulatory configuration at each phoneme boundary against the observed visual configuration. Common failures include incomplete bilabial closure, absent or late labiodental contact, and vowel space compression (reduced visual distinction between open and close vowels).

Understanding how lip-sync analysis is implemented at scale is covered in detail on our how it works page.

Dimension 4: Compression Masking

One of the most significant factors contributing to deepfake believability is compression masking — the phenomenon by which video compression (H.264, H.265, VP9, AV1) destroys the fine-grained artefacts that would otherwise reveal synthesis. This is not an intentional strategy by deepfake creators (though some exploit it deliberately); it is a natural consequence of how video compression works.

Video codecs operate by reducing information in ways that minimise perceptual impact. They discard high-frequency spatial detail (which is where pore-level synthesis artefacts live), smooth temporal variation (which masks texture flickering), and quantise colour information (which obscures subtle chromatic inconsistencies at blend boundaries). The result is that a deepfake that shows clear artefacts at its native render quality may appear artefact-free after a single encoding pass — and social media platforms apply at least one encoding pass, often two or three.

The specific masking effects of each major codec include:

• H.264 — The most widely used codec for social media video. Its 16×16 macroblock structure and DCT-based compression introduce blocking artefacts that can mask the boundary artefacts of face compositing. At lower bitrates (typical of platform re-encoding), the quantisation is aggressive enough to eliminate most spatial-domain synthesis artefacts.
• H.265 (HEVC) — Uses variable-size coding tree units (up to 64×64) and provides better quality at lower bitrates than H.264. Its more sophisticated prediction modes reduce blocking artefacts but introduce their own forensic signatures that can actually help distinguish compressed deepfakes from compressed authentic content — because the codec's prediction models interact differently with synthetic versus authentic content.
• VP9 and AV1 — Used by YouTube and other Google services. These codecs use superblocks and provide competitive compression with H.265. AV1 in particular uses sophisticated in-loop filtering that can smooth over subtle artefacts but also introduces codec-specific signatures that can reveal re-encoding history.

Forensic analysis must account for compression masking by using techniques that are robust to lossy encoding. The most robust approaches operate on temporal patterns (which survive compression better than spatial patterns) and statistical distributions (which are partially preserved even after aggressive quantisation).

Dimension 5: Lighting Consistency

Light in the real world obeys physical laws — it travels in straight lines, reflects according to surface properties, and casts shadows whose geometry is determined by the positions of light sources and occluding objects. A convincing deepfake must produce a face whose shading is consistent with the lighting environment visible in the background.

Modern face-synthesis models handle primary lighting (key light direction and intensity) with reasonable accuracy because the training data implicitly encodes lighting relationships. However, secondary lighting effects are more frequently incorrect:

• Ambient occlusion — The darkening in concavities (under the brow ridge, under the nose, in the nasolabial fold) should be consistent with the ambient light level and direction. Deepfakes often show ambient occlusion that is too uniform or too intense relative to the scene lighting.
• Colour temperature variation — Real-world lighting often involves multiple sources with different colour temperatures (warm key light, cool fill light, coloured environmental reflections). The face should show colour temperature gradients consistent with these sources. Deepfakes frequently show homogeneous colour temperature across the face, failing to model multi-source lighting interaction.
• Environmental reflections — In authentic footage, the eyes and skin surface reflect the surrounding environment. The catch-lights in the eyes should be geometrically consistent with the visible light sources, and skin specular reflections should show colour influence from nearby coloured surfaces. Forensic analysis of eye catch-lights can reveal whether the reflected environment matches the claimed setting.

Dimension 6: Background Integration

The integration between the synthetic face and the background scene is a critical component of believability. Three aspects of background integration determine whether the composite appears seamless:

• Blending boundary quality — The transition between the synthetic face and the original head or background must be invisible. Current techniques use soft alpha masks to blend the regions, but the blend zone produces detectable artefacts in gradient analysis, colour channel separation, and noise profile comparison. The boundary is typically 4–8 pixels wide and follows the face segmentation mask used by the synthesis model.
• Depth coherence — The apparent depth of the synthetic face should be consistent with the surrounding scene. If the original footage shows shallow depth of field (background blur), the synthetic face should exhibit the same blur characteristics. Deepfakes sometimes show a face that is sharper than the background would suggest, or that lacks the expected depth-of-field gradient across the face (ears slightly blurred relative to the nose tip).
• Background stability — As discussed in other investigations, some synthesis models introduce subtle geometric distortions in the background near the face boundary — a "ripple" effect caused by the model's processing of the full frame. This ripple is visible in frame-difference analysis and is one of the most reliable indicators of full-frame synthesis approaches.

Dimension 7: Motion Naturalness

Human motion is characterised by specific kinematic properties that are difficult for generative models to fully reproduce. The most diagnostically useful motion properties are:

• Head motion dynamics — Authentic head motion follows kinematic principles: acceleration and deceleration are smooth, rotation around different axes (yaw, pitch, roll) is coordinated, and the motion exhibits both voluntary (conversational gesturing) and involuntary (vestibular stabilisation) components. Deepfakes often show head motion that is too smooth (lacking the micro- corrections of vestibular stabilisation) or temporally simplified (missing the overlapping voluntary and involuntary components).
• Eye saccade dynamics — Human eyes make rapid movements (saccades) that follow the main sequence — a stereotyped relationship between saccade amplitude and peak velocity. Saccades in deepfakes often violate the main sequence, showing either reduced peak velocity or unnaturally linear velocity profiles (authentic saccades have a characteristic bell-shaped velocity curve).
• Blink dynamics — Authentic blinks follow a specific temporal profile: rapid closure (approximately 80ms), brief hold (approximately 50ms), and slower opening (approximately 150ms). Blink rate is variable (15–20 per minute on average) with distribution influenced by cognitive load, emotional state, and environmental conditions. Deepfakes characteristically show regular blink rates with simplified temporal profiles — symmetric closure and opening rather than the asymmetric authentic profile.

Why Human Perception Misses What Forensic Analysis Catches

The gap between human perception and forensic detection is not accidental — it is a direct consequence of how human vision works. The human visual system is optimised for ecological validity (recognising objects, faces, and threats in natural environments), not for detecting statistical anomalies in pixel data. Several specific properties of human vision contribute to deepfake susceptibility:

• Holistic face processing — Humans process faces holistically rather than analytically. We perceive the overall configuration (the "gestalt") rather than individual features. This means we are good at recognising whether a face looks like a specific person but poor at detecting subtle geometric distortions, texture inconsistencies, or lighting anomalies that do not disrupt the overall facial gestalt.
• Temporal integration — Human vision integrates information across a temporal window of approximately 100–200ms. Frame-to-frame inconsistencies that last less than 2–3 frames are smoothed over by this temporal integration, making texture flickering and micro-expression timing anomalies invisible at playback speed.
• Attention-driven processing — We allocate visual processing resources based on attention, and deepfake content is designed to direct attention to the face and speech content rather than to boundary regions, background stability, or peripheral detail. The forensically relevant artefacts are concentrated precisely in the regions that viewers are least likely to scrutinise.
• Confirmation bias — When content aligns with prior beliefs or expectations, viewers require less evidence to accept it as genuine. Deepfakes are typically designed to be plausible within an existing narrative context, exploiting confirmation bias to lower the bar of perceived authenticity.

This perception gap is why automated forensic analysis is essential. Human reviewers, no matter how experienced, cannot reliably detect state-of-the-art deepfakes through visual inspection alone. The artefacts that distinguish synthetic from authentic content exist at a level of granularity — sub-pixel geometric distortions, statistical noise properties, phoneme-level timing anomalies — that is below the threshold of human perceptual access.

Our forensic analysis platform is designed to operate at precisely this sub-perceptual level, applying statistical analysis, signal processing, and machine learning to detect the artefacts that human vision cannot access. If you have a video you need to verify, our video upload tool provides automated screening across the seven dimensions described in this article. For a complete explanation of our detection methodology, see how our analysis works.