Skip to content

( research · ai video · 27 Oct 2026 )

AI video looks fake when it is too perfect.

Viewers are near chance at spotting a well made AI clip of one person talking. What gives clips away now is polish, like symmetrical waxy faces, studio light, a voice with no room or breath, long shots and anything involving hands or physics. Our videos lose most of their realism to fast settings, small renders and clean output, not to the prompt.

By Sid18 min read57 sources linked
( 01 · summary )

The short version

People are bad at spotting AI video. A meta-analysis of 56 studies put pooled accuracy at 55.5%, and 57.3% for video (Diel et al. 2024). Synthetic faces are judged human more often than real ones and rated 7.7% more trustworthy (Nightingale and Farid, PNAS 2022). The tells that remain are specific. Faces that are too average (symmetrical, smooth, well proportioned), sound that doesn't fit the picture, physics (hands, objects, cutting, pouring) and time (accuracy rises from 72% at 1 second of viewing to 80% at 10). Practitioners who make AI UGC that passes all do the same things: a messy phone-photo first frame, "shot on iPhone, handheld" capture, short lines with ums and breaths, room tone under the voice, clips under 5 seconds, grain and re-compression at the end. Our stack works against several of these. InfiniteTalk runs at 480p with a 4-step speed LoRA that turns guidance off, LTX-2.5 runs distilled and skips its own detail stage, FlashVSR then invents skin, and our voices come out studio-clean. Each section below ends in changes we can make, ranked at the end by gain and cost.

( 02 · studies )

What the studies say gives AI video away

Across the research, people fail on the image and catch fakes on delivery and physics. Groh's study of 2,215 people found the same fakes were caught 58% of the time from a transcript, 62% as silent video, 65% as audio and 74% as video with sound. People judged how something was said more than what was said.

How often people catch a deepfake, by formatAccuracy · 2,215 people · 32 political speech clips · 2024
  1. 01Video with audio74%
  2. 02Audio only65%
  3. 03Silent video with subtitles62%
  4. 04Transcript only58%

Groh et al., Nature Communications 15:7629. 50% is chance.

How often people catch AI images, by viewing timeAccuracy on AI images · 50,000+ people · 750,000 judgments
  1. 0110 seconds80%
  2. 025 seconds77%
  3. 031 second72%

Kamali, Groh et al., CHI 2025. Single portraits were harder to catch than group photos.

Too average reads as AI.

Modern AI faces fail by being symmetrical, well proportioned and typical, not by glitching

EvidenceWith obvious flaws removed, ordinary viewers scored 50.7% and super-recognisers 57.3% (Dunn et al., BJP 2026). People most fooled were most confident (Miller 2023)
What to doAsymmetry, pores, blemishes, flyaway hair, uneven light, ordinary faces. Reject the most symmetrical candidate

Physics and object permanence are the clearest video tell

EvidenceIn Video Reality Test, humans caught fakes 87 to 89% of the time with physical action (a sliced tomato keeping its shape) (arXiv 2512.13281). The best model gets only 22% of hard actions right on VideoPhy-2
What to doKeep hands out of close-ups. No cutting, pouring, handing over or occluding things

Simple scenes hide more

EvidenceSingle portraits were harder to catch than group photos. Each extra person adds gaze, lighting and face-body checks (Hu et al., ICCV 2025)
What to doOne person, one action, plain setting

Sound has to fit the scene

EvidenceAdding audio raised detection only 2.5 to 5 points, and the main audio tell was mismatch, such as speech where room sound belonged (Video Reality Test)
What to doVoice recorded "in" the room the picture shows, ambience under pauses

Compression and lower resolution hide artifacts

EvidenceDetectors lost 29 AUC points on fully generated people video and over 30 more under heavy compression (SynthForensics)
What to doDeliver at 1080p or lower with phone-like compression, not pristine 4K

Motion that is too smooth is a signal

EvidenceGenerated optical flow is smoother and blurrier at edges than real flow (AIGVDet)
What to doSlight handheld drift, small fidgets, real cuts

Suspicion changes how people watch

EvidenceEye tracking shows viewers who suspect AI start hunting for anomalies (Shi 2026)
What to doAvoid anything that triggers the hunt: ad polish, stock-photo faces, perfect light

Uncanny comes from mismatch

Evidence72-study meta-analysis: face distortion is the strongest uncanny trigger, g = 1.46 (Diel, MacDorman 2022)
What to doKeep every part at the same level: a sharp face on a soft body, or a dry voice on a phone clip, breaks it
( 03 · our gap )

Why our videos look less real

Kapwing's people-specific test of Sora 2, Veo 3.1, Kling and Seedance concluded that realism is limited less by pixel quality than by how face, voice, motion and objects hold together over the whole clip. Sora 2 felt most human because of small timing quirks; Veo 3.1 had the best lip sync but looked like a "generated presenter" (Kapwing). Our pipeline adds the presenter look at several steps.

1

Talking-head steps.

InfiniteTalk with the lightx2v 4-step LoRA

What happensText CFG drops to 1 and audio CFG to 2; official settings are 40 steps, text CFG 5, audio CFG 3 to 5 (InfiniteTalk)
Why it costs realismSofter mouths, fewer micro-expressions, more "average" faces. No negative prompt at CFG 1
2

Small renders.

InfiniteTalk at 480p, then upscaled

What happensThe face covers few latent pixels, so pores, teeth, eye wetness and hair strands are never generated
Why it costs realismThe upscaler invents them, which is where waxy or over-crisp skin comes from
3

Upscaler.

FlashVSR on every clip

What happensSmooths and invents detail. A low-denoise Wan 2.2 detail pass refined faces better than FlashVSR or SeedVR2 in one ComfyUI write-up (Sogni)
Why it costs realismPlastic skin next to sharp edges reads as fake
4

LTX-2.5 distilled.

8 steps, CFG 1, FlashVSR after

What happensSkips LTX's own two-stage pipeline: latent x2 upsample, then a refine with a detail IC-LoRA at 1088x1920 (Lightricks)
Why it costs realismThe stage built to add real detail never runs
5

Base model.

LTX-2.5

What happensAbout 150 to 200 Elo behind the leaders on Artificial Analysis, and about 150 behind the best open model, MiniMax H3 (AA)
Why it costs realismNo setting closes all of that gap
6

Locked presenter shot.

Still photo in, audio-driven head out

What happensLocked camera, frozen background, centred subject
Why it costs realismThe "generated presenter" look. InfiniteTalk's README suggests turning the still into a slow pan or zoom clip first
7

First frames.

Clean ChatGPT or FLUX portraits

What happensStudio-even light, smooth skin, symmetrical faces
Why it costs realismThe video model inherits the polish. Practitioners call the first frame "80% of a successful video"
8

Voice.

Chatterbox, Drama 3 or ElevenLabs, dry

What happensNo room, no breath, metronomic pauses
Why it costs realismBreath patterns alone separated real and fake audio almost perfectly in one detector study (Layton et al.)
9

Finish.

Sharp, clean, high-bitrate export

What happensNo grain, no compression, no lens flaws
Why it costs realismPristine footage does not look like a phone

Quantisation is a small factor. Wan's authors report little loss from fp8 weights but real loss from fp8 attention, and plain fp8 on face layers caused grid noise in Wan 2.2 Animate until scaled fp8 was used (arXiv 2503.20314).

( 04 · playbook )

How creators make AI people pass

Collected from AI UGC tools (Arcads, HeyGen, Higgsfield), Google's Veo 3.1 guide, prompt libraries and creators republished from X. Most of it agrees.

35 techniques
A

First frame

techniques 1–8

The still decides most of the result. Make it look like a phone photo before anything moves.

1

Make the still first, reuse it everywhere

HowGenerate, refine and pick the still, then animate. Use the same still for every clip of that person
2

Name the phone and its optics

How"iPhone 11, 26mm equivalent, mild barrel softness at edges", wide lens at f/1.8 to 2.2 so the room stays readable. Heavy bokeh reads as a pro camera
3

Bake in sensor flaws

HowISO noise 500 to 900, light lens smudge, mild JPEG compression, highlights allowed to clip
4

Mixed, one-sided light

HowWindow daylight from one side plus a warm lamp from the other. No studio or "dramatic" light
5

Skin, hair and teeth detail

HowPores, peach fuzz, T-zone shine, tired under-eyes, "no artificial skin smoothing"; flyaway strands; off-white teeth with small gaps, or a closed-mouth smile
6

Asymmetry on purpose

HowAsk for natural facial asymmetry and reject the most symmetrical of about four candidates
Sourcemagicshot
7

Lived-in background, in focus

HowTwo or three specific clutter items (mail, keys, a banana). Too many "flaw" words backfire
8

Pick the image model for flaws

HowNano Banana Pro is better than GPT Image 2 at deliberately unpolished candid looks; GPT Image 2 needs the flaws spelled out. Frame chest-up so the face is big
B

Prompt and motion

techniques 9–16
9

Describe the capture, not the quality

How"Shot entirely on an iPhone, vertical 9:16, front camera at arm's length". Drop "8k, hyperrealistic, cinematic, photorealistic": they trigger stock-photo smoothing
10

Subtle handheld drift

How"Slight natural handheld sway, subtle breathing motion, slight tilt as attention shifts". Called the highest-impact realism cue. Subtle, not shaky. Or prop the phone on a counter and lock the frame
11

Micro-expressions, not emotions

How"A half smile", "slight exhale", "pauses and looks down". "Enthusiastic" adds hand waving
SourceArgil
12

Give the eyes a reason to move

HowA glance off camera, a reaction to a sound, looking at the product
Sourcemagicshot
13

Hands parked

HowCrop head and shoulders or keep hands in "active idle": fingers shift, thumbs rub, gestures only on key words
14

Short negative list

How5 to 8 terms, most important first: "plastic skin, frozen lips, jittery eyes, floating teeth, unnatural blinking, studio lighting". Always add "no subtitles, no captions, no text"
15

Motion prompt is action only

HowRestating hair or clothes in the motion prompt makes the face morph
SourceDXBuilder
16

Real motion source

HowDrive the character with motion from a real person filmed on an iPhone. A cinema-camera source pushes it toward uncanny
Sourcechaseai
C

Voice and script

techniques 17–24

Listeners judge voices by prosody, pause placement, hesitations, breathing and laughter more than by individual sounds (San Segundo, Yamagishi et al.). ElevenLabs clones are now judged real 58% of the time against 62% for real voices (Queen Mary 2025).

17

Speech-to-speech for hero reads

HowA real person reads the script, then the voice is converted. Timing, breaths and emphasis carry over. Same-gender, similar-pitch readers convert best
18

One or two fillers, not ten

HowOne or two "um" per 30 s raise naturalness; more lowers perceived competence and sincerity
19

Breaths at phrase starts

HowUse [sighs] / [exhales] tags or splice real recorded breaths 25 to 30 dB under speech, 150 to 400 ms long. Never at even intervals
20

A bit faster than default

HowAbout 1.05 to 1.1x, pauses of 150 to 400 ms, a longer one only before the key line
SourcePMC 2019
21

Short lines

How5 to 20 words per on-camera clip. Longer text goes to voiceover over B-roll. Fast speech and shouting break lip sync
22

Casual, specific claims

How"Three weeks in and I'm not gatekeeping this anymore." Rewrite marketing copy into how people talk
Sourcecuratedai
23

Several takes, pick phrases

HowRender 3 to 5 takes and cut the best phrases together. Drama 3 varies a lot between takes
Sourceeesel
24

Lock the voice

HowOne voice ID and its settings for a character across every clip. A voice change between cuts is an instant giveaway
D

Sound and lip sync

techniques 25–29
26

Room tone under everything

HowVoice around -12 dB, ambience around -24 dB, running under pauses and cuts. Rotate 4 to 6 beds so loops don't repeat
27

Never let audio lead the lips

HowViewers notice audio about 45 ms early but tolerate up to about 125 to 200 ms late. If unsure, nudge audio one frame later
28

Lip sync on the final audio

HowRun lip sync after the room processing so breaths and pauses drive mouth closures
SourceHighSync on mouth closure in silences
29

Check the mouth on pause

HowTeeth that change shape or merge into a white band, mouth not closing on m/b/p and in silences, a sharper mouth patch than the face. These show on pause even when missed at speed
E

Post and edit

techniques 30–35
30

Clips under 5 seconds

HowIn a 50-person Kling test, AI was spotted 34% of the time under 5 s and 61% over 8 s. Cut every 1 to 2 s, total under about 27 s
Sourcevideoai.me (self-reported), vidau
31

B-roll over the hard parts

HowCover hands, fast phrases and jump cuts with B-roll so the mouth isn't on screen for long stretches
Sourcevidau
32

Order: upscale, blur, grain, grade

HowNever sharpen. Turn LUTs down, lower saturation and contrast to kill the "AI sheen"
33

Grain at 2 to 3%

HowBelow 1% does nothing, above 5% looks like a retro filter
34

1080p, 30 fps, medium bitrate

HowNot 4K, not 60 fps. Some creators re-compress by sending the file through Telegram as a video and posting that copy
35

Leave small imperfections

HowTrim dead air but keep the small stumbles real creators leave in. Shot length matches the line
Sourcecuratedai
( 05 · checklist )

Giveaways to check before posting

Skin

GiveawayWaxy, poreless forehead, cheeks and jaw
FixTextured still, no sharpening, grain (techniques 5, 32, 33)

Teeth

GiveawayOne white block, spacing that shifts while talking
FixOff-white teeth in the still, closed-mouth beats, full-step lip sync

Eyes

GiveawayNo catchlights, odd blink rate, eyes locked on lens
FixA glance off camera (technique 12)

Hair

GiveawaySolid, helmet-like mass with a hard hairline
FixFlyaways and backlight in the still

Hands

GiveawayMerged fingers, objects passing through hands
FixCrop them or park them (technique 13)

Background

GiveawayObjects morphing, garbled signs, frozen room
FixPlain lived-in set, slow pan on the still

Light

GiveawayEven studio light, glossy HDR sheen
FixOne window plus one lamp (technique 4)

Sound

GiveawayDry booth voice, no breaths, silence in pauses
FixRoom chain and room tone (techniques 25, 26)

Sync

GiveawayAudio ahead of lips, long perfectly synced shots
FixAudio never early, short shots (techniques 27, 30)

Consistency

GiveawayFace or voice changing between cuts
FixOne still, one voice ID (techniques 1, 24)

Text

GiveawayHallucinated subtitles or signs
Fix"No subtitles, no captions, no text" (technique 14)

Length

GiveawayOne 15 s unbroken talking shot
FixCuts every 1 to 2 s, B-roll (techniques 30, 31)
( 06 · prompts )

Prompts reported to look real

A first-frame prompt template for a talking head, from a published Nano Banana skill (skills.cat):

A vertical 9:16 UGC-style video frame captured on an iPhone 11 resting on a tripod. Medium-wide portrait
at true eye-level with slightly forward-leaning posture. [CHARACTER] ... [SKIN TEXTURE]: Visible pores
across T-zone, faint smile lines, natural oils catching light on forehead and nose. No filter, no
foundation. [CAMERA]: Native iPhone 11 lens (26mm equivalent), slightly wide perspective, mild barrel
softness at edges. [LIGHTING]: Available light mix: cool overcast daylight from window left, warm
tungsten from desk lamp right. Soft asymmetric shadows, natural falloff. ISO noise 500-900.
[BACKGROUND]: bookshelf, furniture, clearly visible not heavily blurred. [REALITY DETAILS]: Gentle
35mm film grain, light fingerprint smudge on lens, tiny dust haze in air. No cinematic bloom, no
studio finish.

A Kling 3.0 selfie clip from the blind-test article (videoai.me):

Vertical selfie, front-facing camera look, slightly overexposed, natural bathroom light. A woman in
her late 20s, no makeup, hair in a messy bun. 0-2s: adjusts the phone angle. 2-4s: holds up a
skincare tube. 4-5s: says "two weeks and my skin cleared up". Slight lens distortion at edges.
Palette: natural, slightly warm. Negative: studio lighting, perfect skin, frozen lips, professional
framing.

A Veo 3.1 locked-frame testimonial (ugcvids):

9:16 vertical phone-camera shot, one woman in her late twenties in a sunlit bathroom, framed chest-up,
natural window light, no studio lighting. The phone is propped on the counter and the frame is locked:
no pans, no zooms, no dolly. Slightly hurried, slightly amused delivery. She holds a frosted glass
serum bottle up beside her cheek, label facing the lens, holds it still, and says to camera: "Three
weeks. That is it. Three weeks and the texture on my cheeks is gone." Then she sets the bottle down on
the counter. Audio: close conversational voice, faint bathroom room tone, no music. No on-screen text,
no captions, no subtitles, no watermark, no stickers.

The talking_head motion line for InfiniteTalk, with the realism words added:

Handheld front-camera selfie at arm's length, slight natural sway, the woman glances down once before
speaking, small nods on stressed words, a half smile at the end, hands out of frame, background
unchanged, no extra people. Negative: plastic skin, frozen lips, floating teeth, jittery eyes,
studio lighting.
( 07 · audio recipe )

Making a TTS voice sound phone-recorded

Phone mics lose lows, bump the mids and roll off the highs, and a selfie at arm's length picks up the room. This chain was run once on a test clip. Swap the pink noise for 30 s of real room tone recorded on a phone where possible, and swap aecho for a real small-room impulse response (afir) when we have one.

ffmpeg -i voice.wav -f lavfi -i "anoisesrc=color=pink:amplitude=0.004:sample_rate=48000" \
  -filter_complex "[0:a]highpass=f=120,lowpass=f=11000,
    equalizer=f=250:t=q:w=1.2:g=2,equalizer=f=2800:t=q:w=1.5:g=3,equalizer=f=6500:t=q:w=2:g=-4,
    aecho=0.8:0.5:13|29|47:0.28|0.18|0.10,
    acompressor=threshold=-20dB:ratio=3:attack=8:release=120:makeup=3dB,apad=pad_dur=0.4[v];
  [1:a]highpass=f=80,lowpass=f=6000[n];
  [v][n]amix=inputs=2:duration=first:weights=1 1:normalize=0,acrusher=bits=14:mix=0.15,
    alimiter=limit=0.89,loudnorm=I=-14:TP=-1.5:LRA=9,aresample=44100[out]" \
  -map "[out]" -c:a aac -b:a 128k out.m4a
SceneChange
OutdoorsNo reverb, light wind or traffic, low-pass at 9 kHz
Car+2 dB around 150 to 300 Hz, road rumble
Bathroom or kitchenLonger reflections: delays 20 to 60 ms, decays up to 0.35
Music under the voiceDuck it with sidechaincompress, 18 to 25 dB below speech

Last check: listen on a phone speaker and earbuds. If it still sounds like an announcer, lower the 2.8 kHz boost and add a little more room.

( 08 · models )

Video models and settings

Blind preference votes put closed models on top. MiniMax H3 is the best open-weights model on both boards. LTX-2.5, our generator, sits about 150 to 200 Elo below the leaders.

Image-to-video with audio, people's preferenceArtificial Analysis arena Elo · Kling 3.0 Pro anchored at 1000 · read 27 Oct 2026
  1. 01MiniMax H3 768p (open weights)1140
  2. 02MAGI-2 Preview (open, Apache 2.0)1046
  3. 03Kling 3.0 Pro1000
  4. 04LTX-2.5 Fast (ours)984
  5. 05LTX-2.5 Pro958

Blind pairwise votes. Ranks move monthly.

Most natural voicesArtificial Analysis Speech Arena Elo · read 27 Oct 2026
  1. 01ElevenLabs Eleven v41333
  2. 02Qwen-Audio-3.1-TTS-Plus1306
  3. 03Cartesia Sonic 3.61292
  4. 04Fish Audio Drama 3 (open)1245
  5. 05ElevenLabs v31184
  6. 06Chatterbox HD1111

Fish Audio Drama 3 is the best open-weights voice. Chatterbox Turbo, our default, is not in the top 37.

InfiniteTalk

Fast setting (what we run)480p fp8, lightx2v 4 steps, text CFG 1, audio CFG 2
Quality setting40 steps, text CFG 5, audio CFG 3 to 5. Middle ground: lightx2v at 0.75 strength, 8 steps. Test 720p finals

LTX-2.5

Fast setting (what we run)Distilled, 8 steps, CFG 1, FlashVSR after
Quality settingDev about 30 steps, 544x960, latent x2 upsample, refine with detail IC-LoRA at 0.5, 1088x1920

MiniMax H3

Fast setting (what we run)Turbo LoRA, 4 steps
Quality settingFull steps at 768p. Licence reportedly excludes the US, EU, UK and South Korea (NYU Shanghai)

LongCat-Video-Avatar 1.5

Fast setting (what we run)8-step distilled by design
Quality setting720p. MIT licence, Whisper audio encoder. The InfiniteTalk repo points to it as the upgrade. Realism claims are its authors' own (arXiv 2605.26486)

Realism add-ons worth testing: the GalaxyAce phone-footage LoRA for LTX (sensor noise, compression, lens softness, handheld jitter; made for LTX-2 and 2.3, untested on 2.5) (CivArchive) and Instareal for a low-denoise Wan 2.2 detail pass, commercial use allowed (Hugging Face).

( 09 · changes )

Changes for our stack, ranked

Ranked by expected realism gain against cost. None is applied yet. Each should be checked with a blind same-photo, same-voice test before it becomes the default.

Low cost

5
01

Phone-photo first frames.

Nano Banana Pro or GPT Image 2 with the flaw list (techniques 2 to 8), chest-up framing, reject the most symmetrical

Expected gainHigh. The video model inherits whatever polish the still has
02

Room chain on every voice.

The ffmpeg recipe plus real room tone, breath tags, one or two fillers, 1.05x pace

Expected gainHigh. Sound with picture is where viewers catch fakes most (74%)
03

Finish like a phone.

No sharpening, grain 2 to 3%, saturation and contrast down, 1080p 30 fps medium bitrate

Expected gainMedium. Hides artifacts from people and detectors
04

Edit short.

Talking shots under 5 s, cuts every 1 to 2 s, B-roll over hands and fast lines

Expected gainMedium to high. 34% vs 61% spotted in the vendor test
05

Realism prompt defaults

in the Editing agent: capture wording, handheld drift, micro-expressions, short negative list, "no subtitles"

Expected gainMedium. Free

Medium cost

5
06

Full-step talking heads for finals.

InfiniteTalk at 40 steps with CFG, or lightx2v 0.75 at 8 steps. Keep 4 steps for drafts

Expected gainHigh. Sharper mouths and more expression. About 5 to 10x GPU time
07

Test LongCat-Video-Avatar 1.5 at 720p

against InfiniteTalk 480p and 720p

Expected gainHigh if it wins. Same speed class, MIT licence
08

LTX-2.5 two-stage pipeline

instead of distilled plus FlashVSR

Expected gainMedium to high. Real detail instead of invented detail
09

Swap or tame FlashVSR.

Try a low-denoise Wan 2.2 pass with Instareal, upscale only 1.5 to 2x

Expected gainMedium. Less plastic skin
10

Speech-to-speech for hero clips.

A person reads, the voice is converted

Expected gainHigh for testimonials. Timing and breaths are real

High cost

2
11

Pilot MiniMax H3 or MAGI-2

as the main generator

Expected gainLargest raw gain on the arenas. H3's licence needs checking first
12

Pay per clip for a closed model

(Seedance 2.5, Wan 3.0, HeyGen Video 1) on hero shots, our own voice on top

Expected gainHighest quality per clip
( 10 · conclusion )

Conclusion

Today

Clean, fast and polished, which is what reads as AI.

  • Studio-clean first frames
  • 4-step InfiniteTalk at 480p, FlashVSR on top
  • Dry TTS voice with no room or breath
  • Long talking shots, sharp clean export

Next

Imperfect in the places real phone footage is imperfect.

  • Phone-photo stills with flaws (technique 2)
  • Full-step talking heads for finals, LongCat test
  • Room chain, breaths, one or two ums (technique 25)
  • Clips under 5 s, grain, 1080p (technique 30)

( more from the lab )

Keep reading.

We research problems before we build apps for them. Every source is linked, so check our work.