On the standard SyncNet sync test, the VASA-1 talking-face model scores 8.841, higher than the 7.640 earned by the real videos it is compared against (Microsoft Research Asia, VASA-1 paper 2024). That one number explains most of this field in 2026: models now beat real footage on the metric built to judge them, so user studies and runtime figures matter more than ever. A real-time telepresence face model wins 84.59% of A/B comparisons against the slower DiffPoseTalk (Meta and Seoul National University, SIGGRAPH Asia 2025), and 3D training data has grown from typically under 1 hour to 18.5 hours (UniTalker paper, arXiv 2024). Meanwhile, human viewers detect deepfakes at 55.54%, barely above a coin flip (Diel et al., Computers in Human Behavior Reports 2024). We aggregated data from NVIDIA Research, ByteDance, the European Broadcasting Union, Unity, and a dozen benchmark papers from university and industry labs. If you came here from the recognition side of the problem, our lip-reading visual speech recognition statistics cover the reverse task: reading words from lips instead of generating lips from words.
TL;DR
- VASA-1 scores 8.841 SyncNet confidence on VoxCeleb2 vs 7.640 for real video (Microsoft Research Asia, VASA-1 paper 2024).
- LatentSync cuts HDTF FVD to 162.74 from Wav2Lip’s 304.35 (ByteDance, LatentSync paper 2025).
- Lip vertex error on BIWI-Test-A fell from 6.5563 (VOCA) to 4.3646 (TalkingStyle) x 10^-4 mm (GLDiTalker paper, arXiv 2025).
- UniTalker scaled 3D facial training data from typically under 1 hour to 18.5 hours (UniTalker paper, arXiv 2024).
- 92.47% of A/B judgments preferred CodeTalker over VOCA for BIWI lip sync (CodeTalker paper, arXiv 2023).
- A real-time telepresence face model runs in 10 ms of GPU time vs 1.3 s for Audio2Photoreal (Meta and Seoul National University, SIGGRAPH Asia 2025).
- Audio2Face-3D v3.0 reaches 3,269 FPS on one RTX 4090 with TensorRT (NVIDIA, Audio2Face-3D paper 2025).
- 36% of game professionals use generative AI tools, and 52% say it harms the industry (GDC, 2026 State of the Game Industry).
- Only about 12.4% of on-screen dubbed speech time keeps the same viseme as the original (AWS AI Labs and MIT, Dubbing in Practice, TACL 2023).
- Broadcast lip sync must stay within 40 ms early and 60 ms late end to end (EBU Recommendation R37-2007).
- Human deepfake detection averages 55.54% (95% CI 48.87 to 62.10) across 86,155 participants (Diel et al., Computers in Human Behavior Reports 2024).
- Open-source deepfake video detectors lose 50% of their AUC on in-the-wild 2024 content (Deepfake-Eval-2024 benchmark, 2025).
1. Lip-Sync Accuracy Benchmarks for Talking-Head Video
Two families of numbers dominate 2D lip-sync papers. SyncNet confidence (reported as LSE-C or Sync_conf, higher is better) and SyncNet distance (LSE-D, lower is better) come from a network trained to judge whether mouth movement matches audio. FID and FVD (Frechet Inception Distance and Frechet Video Distance, lower is better) measure how far generated frames and clips drift from real ones.
The headline is a paradox. VASA-1 posts a SyncNet confidence of 8.841 on VoxCeleb2, while the real videos score 7.640 (Microsoft Research Asia, VASA-1 paper 2024). A model that is ‘more synced than reality’ on this test is not more realistic: its FVD25 is 105.88 against 29.25 for real video. Sync is close to solved by the metric; visual realism is not. LatentSync shows the other lever: removing SyncNet supervision drops its sync confidence from 8.9 to 4.6, matching the authors’ finding that latent diffusion models pick up visual shortcuts instead of audio-lip links unless a sync judge is built into training (ByteDance, LatentSync paper 2025).
| Metric | Value | Source |
|---|---|---|
| SyncNet confidence on VoxCeleb2 test set: VASA-1 vs real video (higher is better) | 8.841 vs 7.640 | Microsoft Research Asia, VASA-1 paper 2024 |
| FVD25 on OneMin-32 set: VASA-1 vs real video vs SadTalker (lower is better) | 105.88 vs 29.25 vs 214.51 | Microsoft Research Asia, VASA-1 paper 2024 |
| SyncNet confidence (Sync_conf) on HDTF: LatentSync vs Wav2Lip vs MuseTalk | 8.9 vs 8.2 vs 6.8 | ByteDance, LatentSync paper 2025 |
| FVD on HDTF: LatentSync vs Wav2Lip (lower is better) | 162.74 vs 304.35 | ByteDance, LatentSync paper 2025 |
| FID on HDTF: LatentSync vs Wav2Lip (lower is better) | 7.22 vs 12.5 | ByteDance, LatentSync paper 2025 |
| Sync_conf on HDTF: LatentSync without SyncNet supervision vs with pixel-space SyncNet | 4.6 vs 8.9 | ByteDance, LatentSync paper 2025 |
| SyncNet judge accuracy on HDTF test set: StableSyncNet vs previous best | 94% vs 91% | ByteDance, LatentSync paper 2025 |
Context note: Wav2Lip, the 2020 model that popularized SyncNet-based training (Wav2Lip paper, ACM Multimedia 2020), still appears as a baseline in nearly every 2025 table. Its scores move between papers: LatentSync reports a Sync_conf of 8.2 on HDTF, while a separate evaluation reports an LSE-C of 9.054 on HDTF (see section 3). Benchmark numbers are only comparable inside one paper’s evaluation setup.
2. Speech-Driven 3D Faces: Lip Vertex Error and Training Data
For 3D faces (meshes used in games, VR and film), the standard metric is lip vertex error (LVE): the largest per-frame distance between generated and captured lip vertices, averaged over frames. Lower is better. On the BIWI-Test-A benchmark, LVE fell from 6.5563 for VOCA to 4.3646 for TalkingStyle (x 10^-4 mm), a 33% reduction by our math (2.1917 divided by 6.5563) across several model generations (GLDiTalker paper, arXiv revised 2025).
The bigger story is data starvation. The two benchmark datasets the field trains on are tiny: VOCASET holds 480 sequences of about 4 seconds each, and BIWI has 14 speakers reading 40 sentences (CodeTalker paper, arXiv 2023). UniTalker attacked that bottleneck by merging eight datasets and scaling training data to 18.5 hours, which cut LVE by 9.2% on BIWI and 13.7% on VOCASET (UniTalker paper, arXiv 2024). Gains of that size from data alone suggest the next jumps will come from capture budgets, not architecture.
| Metric | Value | Source |
|---|---|---|
| Lip vertex error on BIWI-Test-A: VOCA vs FaceFormer (x 10^-4 mm, lower is better) | 6.5563 vs 5.3077 | GLDiTalker paper, arXiv 2025 (Table 1) |
| Lip vertex error on BIWI-Test-A: CodeTalker vs FaceDiffuser (x 10^-4 mm) | 4.7914 vs 4.7823 | GLDiTalker paper, arXiv 2025 (Table 1) |
| Lip vertex error on BIWI-Test-A: TalkingStyle vs GLDiTalker (x 10^-4 mm) | 4.3646 vs 4.6440 | GLDiTalker paper, arXiv 2025 (Table 1) |
| 3D facial training data: commonly used datasets vs UniTalker | typically under 1 hour vs 18.5 hours | UniTalker paper, arXiv 2024 |
| Lip vertex error reduction from UniTalker’s unified training: BIWI vs VOCASET | 9.2% vs 13.7% | UniTalker paper, arXiv 2024 |
| Average error reduction across the A2F-Bench benchmark | 6.3% | UniTalker paper, arXiv 2024 |
| VOCASET size: sequences, speakers, capture rate, mesh vertices | 480 sequences, 12 speakers, 60 fps, 5,023 vertices | CodeTalker paper, arXiv 2023 |
| BIWI size: speakers, sentences, capture rate, mesh vertices | 14 speakers, 40 sentences, 25 fps, 23,370 vertices | CodeTalker paper, arXiv 2023 |
Context note: GLDiTalker is the only model in that table built to produce clearly varied outputs (diversity score 8.2176, vs 0 for most rivals), a design choice that works against matching one captured lip track exactly. Most recent available data for the two dataset descriptions: a 2023 paper.
3. User-Study Preference Rates
Because LVE punishes any mouth motion that differs from the one captured, even when the alternative looks right, papers lean on human raters. The pattern across studies is consistent: new models crush the oldest baselines but often beat the previous generation by thin margins. CodeTalker won 92.47% of BIWI lip-sync judgments against VOCA but only 53.76% against FaceFormer, near a coin flip (CodeTalker paper, arXiv 2023).
Real-time systems are now winning these tests too. A causal telepresence model, which uses no future audio, was preferred in 84.59% of A/B comparisons against DiffPoseTalk and 69.17% against Audio2Photoreal (Meta and Seoul National University, SIGGRAPH Asia 2025). And the gap between metric and perception is measurable: in a Karlsruhe Institute of Technology study, Wav2Lip had the best HDTF LSE-C of any method yet a human sync rating of only 2.91, against 3.92 for the authors’ model (Karlsruhe Institute of Technology, AV-HuBERT expert paper 2024).
| Metric | Value | Source |
|---|---|---|
| A/B judgments preferring CodeTalker over VOCA for lip sync on BIWI-Test-B | 92.47% | CodeTalker paper, arXiv 2023 (31 participants) |
| A/B judgments preferring CodeTalker over FaceFormer for lip sync on BIWI-Test-B | 53.76% | CodeTalker paper, arXiv 2023 |
| A/B preference for the real-time telepresence face model over DiffPoseTalk | 84.59% | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| A/B preference for the real-time telepresence face model over Audio2Photoreal | 69.17% | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| Mean opinion score for lip sync on VOCASET (1 to 5): GLDiTalker vs FaceFormer | 4.05 vs 3.08 | GLDiTalker paper, arXiv 2025 (20 participants) |
| Human sync rating on HDTF: best model vs Wav2Lip, while Wav2Lip had the top LSE-C of 9.054 | 3.92 vs 2.91 | Karlsruhe Institute of Technology, AV-HuBERT expert paper 2024 (10 participants) |
Context note: sample sizes are small. The studies above used 10, 20 and 31 raters, which is typical for the field and a reason to treat single-digit preference gaps with caution. CodeTalker’s study design produced 372 judgment entries per dataset from 120 sample pairs.
4. Runtime Performance and Latency
Speed has moved from the bottleneck to a solved problem for 3D rigs. The open-source Audio2Face-3D v3.0 diffusion network runs at 3,269 frames per second on a single GeForce RTX 4090 with TensorRT (NVIDIA, Audio2Face-3D paper 2025). The trick is distillation: the model trains with 1,000 diffusion steps but infers with 2. For studios, the practical figure is concurrency: the older v2.3 regression network drives 128 simultaneous animation tracks (one per speaking character) in 1.3 GB of GPU memory.
Photoreal video is still far heavier. VASA-1 streams 512x512 video at up to 40 FPS after a 170 ms starting delay on an RTX 4090 (Microsoft Research Asia, VASA-1 paper 2024), and MuseTalk delivers 30 FPS at only 256x256 on a V100 data-center GPU (MuseTalk paper, arXiv 2024). The latency race mirrors what happened in speech synthesis, covered in our TTS real-time factor and inference cost statistics.
| Metric | Value | Source |
|---|---|---|
| Audio2Face-3D v3.0 diffusion network speed, 1 track, TensorRT, RTX 4090 | 3,269 frames per second | NVIDIA, Audio2Face-3D paper 2025 |
| Audio2Face-3D v2.3 (Claire) speed, TensorRT, RTX 4090: 1 track vs 8 concurrent tracks | 453 vs 413 frames per second | NVIDIA, Audio2Face-3D paper 2025 |
| Audio2Face-3D v2.3 GPU memory with TensorRT: 8 tracks vs 128 tracks | 0.6 GB vs 1.3 GB | NVIDIA, Audio2Face-3D paper 2025 |
| VASA-1 online generation at 512x512 on one RTX 4090: frame rate and starting latency | up to 40 FPS, 170 ms | Microsoft Research Asia, VASA-1 paper 2024 |
| Real-time telepresence face model GPU time per frame on an A100 | 10 ms (100 FPS), with zero audio lookahead | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| Offline baselines in the same comparison: Audio2Photoreal vs DiffPoseTalk with style | 1.3 s vs 11.4 s | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| MuseTalk lip-sync video output on a V100 GPU | 30 FPS at 256x256 | MuseTalk paper, arXiv 2024 |
Context note: the telepresence model is 130 times faster than Audio2Photoreal by our math (1,300 ms divided by 10 ms), consistent with the paper’s own claim of 100 to 1,000 times faster than offline baselines. Audio2Face-3D FPS counts network output frames, not a rendered game scene.
5. Adoption in Games and Sentiment Among Developers
Audio-driven animation is shipping, but into an industry that has turned sharply against generative AI. 52% of game professionals now say generative AI is harming the industry, up from 30% a year earlier and 18% the year before (GDC, 2026 State of the Game Industry). The most negative group is visual and technical art at 64%, the same people whose facial animation work these tools automate.
Usage figures diverge depending on who answers. The survey of individual professionals above finds 36% use generative AI, while a studio-level survey reports 96% of studios use AI tools in select workflows (Unity, 2025 Gaming Report). On the supply side, the Audio2Face models, SDK and training framework went open source on September 24, 2025, with nine companies named as already integrating them (NVIDIA Developer Blog 2025). The voice half of this pipeline is tracked in our AI voice in game NPC dialogue adoption statistics.
| Metric | Value | Source |
|---|---|---|
| Game industry professionals using generative AI tools in their job | 36% (of more than 2,300 respondents) | GDC, 2026 State of the Game Industry |
| Generative AI use: respondents at game studios vs at publishers, support and marketing/PR firms | 30% vs 58% | GDC, 2026 State of the Game Industry |
| Respondents saying generative AI has a negative impact on the industry: 2026 vs prior year vs two years earlier | 52% vs 30% vs 18% | GDC, 2026 State of the Game Industry |
| Visual and technical art professionals with an unfavorable view of generative AI | 64% | GDC, 2026 State of the Game Industry |
| Respondents saying generative AI has a positive impact: 2026 vs 2025 | 7% vs 13% | GDC, 2026 State of the Game Industry |
| Studios using AI tools in select workflows | 96% | Unity, 2025 Gaming Report |
| Respondents feeling positive about AI tools in game development | 79% | Unity, 2025 Gaming Report |
| Companies named as Audio2Face adopters at the open-source launch (vendor’s own claim) | 9 companies, plus 2 shipped titles credited | NVIDIA Developer Blog, September 24, 2025 |
Context note: the nine named adopters are Convai, Codemasters, GSC Games World, Inworld AI, NetEase, Reallusion, Perfect World Games, Streamlabs and UneeQ; the two titles are Alien: Rogue Incursion Evolved Edition (Survios) and Chernobylite 2: Exclusion Zone (The Farm 51). Neither studio quote attached a number to the time saved.
6. Dubbing Reality and Broadcast Sync Tolerances
AI visual dubbing promises to make translated speech match the actor’s mouth. The baseline it competes with is looser than most people assume. In 319.57 hours of professionally dubbed English-to-German and English-to-Spanish TV and film, only about 12.4% of on-screen speech time had the same viseme in the dub as in the original (AWS AI Labs and MIT, Dubbing in Practice, TACL 2023). A viseme is the visible mouth shape for a group of sounds (for example, p, b and m all close the lips). Human dubbers match mouth shapes somewhat more often when the speaker is on screen, yet the authors conclude that vocal naturalness and translation quality matter more to professionals than strict lip sync.
Timing is the tighter constraint. Isochrony (the dub’s speech starting and stopping when the original speech does) averaged a 0.658 overlap. Broadcasters, for their part, set hard limits on audio-video offset: sound may lead picture by at most 40 ms and lag by at most 60 ms, values set where the mismatch becomes perceptible to 50% of observers (EBU Recommendation R37-2007). Any live avatar pipeline inherits that budget, so a generator’s starting latency (170 ms for VASA-1, see section 4) has to be absorbed by delaying the audio to match, or the result fails a broadcast check.
| Metric | Value | Source |
|---|---|---|
| Professionally dubbed corpus analyzed: hours, titles, episodes | 319.57 hours, 54 titles, 674 episodes | AWS AI Labs and MIT, Dubbing in Practice, TACL 2023 |
| Share of on-screen speech time with the same viseme in source and dub | about 12.4% | AWS AI Labs and MIT, Dubbing in Practice, TACL 2023 |
| Mean (median) speech-time overlap between original line and dubbed line | 0.658 (0.731) | AWS AI Labs and MIT, Dubbing in Practice, TACL 2023 |
| Speech-time overlap for on-screen speakers vs off-screen speakers | 0.684 vs 0.662 | AWS AI Labs and MIT, Dubbing in Practice, TACL 2023 |
| Dubbed lines at least 10% longer in characters than the English source: German vs Spanish | 53% vs 30% | AWS AI Labs and MIT, Dubbing in Practice, TACL 2023 |
| End-to-end broadcast limit for sound relative to picture: sound early vs sound late | at most 40 ms vs at most 60 ms | EBU Recommendation R37-2007 |
| Sync accuracy required at each production stage: sound early vs sound late | 5 ms vs 15 ms | EBU Recommendation R37-2007 |
Context note: most recent available data: the dubbing corpus study, 2023, and EBU R37, 2007 (still the current EBU recommendation). The dubbing study also found that 4.3% of dubbed lines had zero time overlap with the original, and that translation quality scores did not differ significantly between on-screen and off-screen lines (COMET-QE p = 0.349).
7. Perception and the Deepfake-Detection Angle
The same models that animate game characters can fake a real person’s speech. Detection is lopsided. Across 56 papers and 86,155 participants, people detected deepfakes with 55.54% accuracy (95% CI 48.87 to 62.10), not significantly above chance, and only 57.31% for video (Diel et al., Computers in Human Behavior Reports 2024). Feedback training, AI support or caricaturization lifted accuracy to 65.14%.
Machines specialized for lip sync do much better in the lab, but generalization is the weak spot. The LipFD detector, which hunts for timing mismatches between audio and lip movement, reports more than 95.3% average accuracy and up to 90.2% in real-world settings such as WeChat video calls (Lips Are Lying, LipFD paper, arXiv 2024). On deepfakes actually circulating in 2024, open-source video detectors lost 50% of their AUC (area under the ROC curve, where 1.0 is perfect and 0.5 is chance) compared with older benchmarks (Deepfake-Eval-2024 benchmark, 2025).
| Metric | Value | Source |
|---|---|---|
| Pooled human deepfake detection accuracy, all media types | 55.54% (95% CI 48.87 to 62.10, k = 67) | Diel et al., Computers in Human Behavior Reports 2024 |
| Human detection accuracy for deepfake video specifically | 57.31% (95% CI 47.80 to 66.57, k = 26) | Diel et al., Computers in Human Behavior Reports 2024 |
| Human detection accuracy with feedback training, AI support or caricaturization | 65.14% (95% CI 55.21 to 74.46, k = 15) | Diel et al., Computers in Human Behavior Reports 2024 |
| LipFD lip-sync deepfake detector: average accuracy vs real-world scenarios | more than 95.3% vs up to 90.2% | Lips Are Lying (LipFD) paper, arXiv 2024 |
| AUC drop of open-source detectors on 2024 in-the-wild deepfakes vs earlier benchmarks: video, audio, image | 50%, 48%, 45% | Deepfake-Eval-2024 benchmark, 2025 |
| Deepfake-Eval-2024 dataset: video hours, websites, languages | 45 hours, 88 websites, 52 languages | Deepfake-Eval-2024 benchmark, 2025 |
| Accuracy of the VASA-1 authors’ own detector at telling real videos from VASA-1 output | 97.8% | Microsoft Research Asia, VASA-1 paper 2024 |
Context note: the 97.8% figure covers one model’s output with a detector trained for it. The Deepfake-Eval-2024 result shows what happens when detectors meet generators they were not trained on.
Summary: Audio-Driven Facial Animation and Lip-Sync by the Numbers
| Metric | Value | Source |
|---|---|---|
| SyncNet confidence on VoxCeleb2: VASA-1 vs real video | 8.841 vs 7.640 | Microsoft Research Asia, VASA-1 paper 2024 |
| FVD on HDTF: LatentSync vs Wav2Lip | 162.74 vs 304.35 | ByteDance, LatentSync paper 2025 |
| SyncNet judge accuracy on HDTF: StableSyncNet vs previous best | 94% vs 91% | ByteDance, LatentSync paper 2025 |
| Lip vertex error on BIWI-Test-A: VOCA vs TalkingStyle (x 10^-4 mm) | 6.5563 vs 4.3646 | GLDiTalker paper, arXiv 2025 |
| 3D facial training data: typical datasets vs UniTalker | under 1 hour vs 18.5 hours | UniTalker paper, arXiv 2024 |
| A/B judgments preferring CodeTalker over VOCA for lip sync on BIWI | 92.47% | CodeTalker paper, arXiv 2023 |
| A/B preference for the real-time telepresence face model over DiffPoseTalk | 84.59% | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| Human sync rating on HDTF: best model vs Wav2Lip despite Wav2Lip’s top LSE-C | 3.92 vs 2.91 | Karlsruhe Institute of Technology, 2024 |
| Audio2Face-3D v3.0 speed on one RTX 4090 with TensorRT | 3,269 frames per second | NVIDIA, Audio2Face-3D paper 2025 |
| Real-time telepresence face model GPU time per frame | 10 ms | Meta and Seoul National University, SIGGRAPH Asia 2025 |
| Game professionals using generative AI tools | 36% | GDC, 2026 State of the Game Industry |
| Studios using AI tools in select workflows | 96% | Unity, 2025 Gaming Report |
| On-screen dubbed speech time with the same viseme as the original | about 12.4% | AWS AI Labs and MIT, TACL 2023 |
| Mean speech-time overlap between original and dubbed lines | 0.658 | AWS AI Labs and MIT, TACL 2023 |
| Broadcast end-to-end lip-sync limit: sound early vs sound late | 40 ms vs 60 ms | EBU Recommendation R37-2007 |
| Pooled human deepfake detection accuracy | 55.54% | Diel et al., Computers in Human Behavior Reports 2024 |
| AUC drop of open-source video deepfake detectors on 2024 in-the-wild content | 50% | Deepfake-Eval-2024 benchmark, 2025 |
Methodology and Sources
- NVIDIA: Audio2Face-3D, Audio-driven Realistic Facial Animation for Digital Avatars (arXiv 2508.16401, 2025) for inference speed, memory and diffusion-step figures, and the Audio2Face open-source announcement, September 24, 2025 for named adopters (vendor’s own claim).
- Microsoft Research Asia, VASA-1 paper (arXiv 2404.10667, revised 2024): sync scores, FVD, frame rate, latency, detector accuracy.
- Meta Codec Avatars Lab and Seoul National University, Audio Driven Real-Time Facial Animation for Social Telepresence (SIGGRAPH Asia 2025): latency and A/B preference.
- ByteDance, Beijing Jiaotong University and Monash University, LatentSync (arXiv 2412.09262, revised 2025): HDTF and VoxCeleb2 sync, FID and FVD.
- GLDiTalker paper, Beihang University and others (arXiv 2408.01826, revised 2025): BIWI lip vertex error table and mean opinion scores.
- CodeTalker paper (arXiv 2301.02379, 2023): A/B preference rates and dataset descriptions.
- UniTalker paper (arXiv 2408.00762, 2024): training-data scale and error reductions.
- Karlsruhe Institute of Technology, Istanbul Technical University and Carnegie Mellon University, AV-HuBERT expert paper (arXiv 2405.04327, 2024): LSE-C/LSE-D vs user study.
- MuseTalk paper (arXiv 2410.10122, revised 2025): real-time frame rate.
- Wav2Lip paper (ACM Multimedia 2020): background on the baseline model, no numbers cited from it directly.
- GDC, 2026 State of the Game Industry: generative AI use and sentiment.
- Unity, 2025 Gaming Report launch post (March 17, 2025): studio AI use and sentiment.
- AWS AI Labs and MIT Media Lab, Dubbing in Practice (TACL vol. 11, 2023): professional dubbing corpus statistics.
- EBU Recommendation R37-2007: sound and vision timing limits.
- Diel et al., Human performance in detecting deepfakes, Computers in Human Behavior Reports vol. 16 (2024): meta-analysis of human detection.
- Lips Are Lying (LipFD) paper (arXiv 2401.15668, 2024): lip-sync deepfake detection accuracy.
- Deepfake-Eval-2024 benchmark (arXiv 2503.02857, 2025): in-the-wild detector performance.
- Data watch: (1) Metric-vs-perception conflicts are real: VASA-1 beats real video on SyncNet confidence, and Wav2Lip’s HDTF score differs between papers (Sync_conf 8.2 in the LatentSync evaluation vs LSE-C 9.054 in the Karlsruhe evaluation), so compare scores only within one paper. (2) The 36% and 96% AI-use figures in section 5 measure different populations (individual professionals vs studios, and the studio survey’s sample size is not stated on its launch page), so they should not be read as a contradiction. (3) We found no user-study results in the Audio2Face-3D paper text we read, so its perceptual quality is not covered here. (4) CodeTalker (2023), the dubbing corpus study (2023) and EBU R37 (2007) are the most recent available data for their metrics; newer SIGGRAPH and CVPR papers may update the LVE and preference figures. (5) LVE units differ across benchmarks and papers, so values are only comparable within the same table. (6) User studies in this field rely on 10 to 31 raters, so small preference gaps are not robust.
Last updated: November 6, 2026. We update this roundup quarterly, and the next refresh is expected when the GDC 2027 State of the Game Industry report is published.