| Variable | Coefficient (β) | p‑value | Interpretation | |----------|-----------------|---------|----------------| | Video length (seconds) | 0.0045 | <0.001 | Each additional second adds ≈ 0.45 % to price. | | **Resolution (4K vs. 1080p
| Year | Model / System | Core Architecture | Notable Metrics (on standard benchmark) | |------|----------------|-------------------|----------------------------------------| | 2018 | Face2Face | Real‑time facial reenactment (3‑D morphable models) | 85 % SSIM, 73 % user‑perceived realism | | 2019 | DeepFakeLab | Encoder‑decoder GAN + facial landmarks | 88 % SSIM, 71 % user‑perceived realism | | 2020 | First Order Motion Model (FOMM) | Keypoint‑based motion transfer | 91 % LPIPS, 75 % user‑perceived realism | | 2021 | StyleGAN‑Video | Temporal StyleGAN with latent interpolation | 93 % LPIPS, 78 % user‑perceived realism | | 2022 | RunwayGen‑2 | Text‑to‑video diffusion (unconditional) | 94 % FVD, 80 % user‑perceived realism | | 2023 | DeepFaceLive | Real‑time GAN + audio‑driven lip sync | 95 % LPIPS, 82 % user‑perceived realism | | 2024 | Fantopiamond (focus of this paper) | Dual‑latent diffusion + Temporal Consistency Transformer + Audio‑Conditioned Lip‑Sync | 97 % LPIPS, 88 % human Turing‑test | fantopiamondomongerdeepfakesmargotrobbiea top
| Metric | Description | Target | |--------|-------------|--------| | LPIPS (Learned Perceptual Image Patch Similarity) | Perceptual similarity (lower = better) | ≤0.05 | | FVD (Fréchet Video Distance) | Distributional distance between real and generated video | ≤30 | | Human Turing‑Test | % of participants who mistake fake for real after a 30‑second view | ≥85 % | | Temporal Flicker Index | Standard deviation of pixel differences across adjacent frames | ≤0.02 | | Audio‑Visual Sync Score | Cross‑modal correlation between phoneme onset and lip closure | ≥0.93 | | Variable | Coefficient (β) | p‑value |