Too average reads as AI.
Modern AI faces fail by being symmetrical, well proportioned and typical, not by glitching
( research · ai video · 27 Oct 2026 )
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.
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.
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.
Groh et al., Nature Communications 15:7629. 50% is chance.
Kamali, Groh et al., CHI 2025. Single portraits were harder to catch than group photos.
Modern AI faces fail by being symmetrical, well proportioned and typical, not by glitching
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.
InfiniteTalk with the lightx2v 4-step LoRA
InfiniteTalk at 480p, then upscaled
FlashVSR on every clip
8 steps, CFG 1, FlashVSR after
LTX-2.5
Still photo in, audio-driven head out
Clean ChatGPT or FLUX portraits
Chatterbox, Drama 3 or ElevenLabs, dry
Sharp, clean, high-bitrate export
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).
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.
The still decides most of the result. Make it look like a phone photo before anything moves.
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).
Nothing matches. Try another word or clear the filters.
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.
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| Scene | Change |
|---|---|
| Outdoors | No reverb, light wind or traffic, low-pass at 9 kHz |
| Car | +2 dB around 150 to 300 Hz, road rumble |
| Bathroom or kitchen | Longer reflections: delays 20 to 60 ms, decays up to 0.35 |
| Music under the voice | Duck 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.
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.
Blind pairwise votes. Ranks move monthly.
Fish Audio Drama 3 is the best open-weights voice. Chatterbox Turbo, our default, is not in the top 37.
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).
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.
Nano Banana Pro or GPT Image 2 with the flaw list (techniques 2 to 8), chest-up framing, reject the most symmetrical
The ffmpeg recipe plus real room tone, breath tags, one or two fillers, 1.05x pace
No sharpening, grain 2 to 3%, saturation and contrast down, 1080p 30 fps medium bitrate
Talking shots under 5 s, cuts every 1 to 2 s, B-roll over hands and fast lines
in the Editing agent: capture wording, handheld drift, micro-expressions, short negative list, "no subtitles"
InfiniteTalk at 40 steps with CFG, or lightx2v 0.75 at 8 steps. Keep 4 steps for drafts
against InfiniteTalk 480p and 720p
instead of distilled plus FlashVSR
Try a low-denoise Wan 2.2 pass with Instareal, upscale only 1.5 to 2x
A person reads, the voice is converted
as the main generator
(Seedance 2.5, Wan 3.0, HeyGen Video 1) on hero shots, our own voice on top
Clean, fast and polished, which is what reads as AI.
( more from the lab )
We research problems before we build apps for them. Every source is linked, so check our work.