A reference-conditioned animator ships with a graph feeding its face condition port a perframe crop of the driver’s face that suppresses the target identity rather than transferring the driver’s, and a portrait animator outperforms everything measured here.
Abstract
A reference-conditioned animator ships with a graph feeding its face condition port a perframe crop of the driver’s face. That port suppresses the target identity rather than transferring the driver’s. We audit it on eight identities, three motion regimes from an optical-flow index, two drivers and two seeds: 96 cells per configuration. Re-pointing it at the reference photograph at half weight raises ArcFace similarity to the reference by a median + 0.1164 on 94 of 96 cells, 2.67 noise floors; similarity to the driver averages −0.001 to + 0.020, inside every impostor band. The port is one identity–expression knob: attenuation is cheap, freezing its content carries the entire expression cost on two calipers, and swapping who it shows buys identity at no measurable cost. That predicts a working point: a reference face animated by the driver gains a mean + 0.0922 over the stock wiring (89 of 96 cells, 2.13 noise floors) with no separable difference on the geometric mouth caliper (<inline-formula> <tex-math notation="LaTeX">$p = 0.377$ </tex-math></inline-formula>). Eighteen viewers prefer the stock wiring to the frozen port in 164 of 216 naturalness comparisons; identity-match ratings do not separate the four arms (<inline-formula> <tex-math notation="LaTeX">$p = 0.89$ </tex-math></inline-formula>). Four results are negative: a cartoon intermediate buys no identity (0.36 noise floors, <inline-formula> <tex-math notation="LaTeX">$p = 0.093$ </tex-math></inline-formula>) at <inline-formula> <tex-math notation="LaTeX">$1.82\times $ </tex-math></inline-formula> the compute and a 5.2% failure rate; an edit made there does not survive restoration; a face swap wins the identity metric while increasing perceptual distance (92 of 94 cells). For face-dominant content a portrait animator outperforms everything measured here.
AlbumVSR conditions a frozen, one-step streaming diffusion transformer on a personal album, which uses photographs of the same person as an external source of identity.
Nian-Yao Du, Xi-Jun Wang, Stanley H. Chan· 0 citations
Thermal-to-visible (T2V) face translation has important applications in nighttime biometric recognition, low-light surveillance, and human interpretation of thermal observations. The task remains challenging because thermal images do not directly record visible appearance, while the cross-spectral cues useful for recon...
Hao-Jie Li, Xin-Long Sun, Xiaohaoyang Lei et al.· 0 citations
Identity-preserving video generation aims to maintain a subject's identity while synthesizing realistic videos. Yet a single reference portrait captures the subject's appearance under only one facial configuration. As expressions change, facial appearance can vary in highly identity-specific ways, leaving the subject's...
Tian-Wen Fu, Wen-Bin Teng, Gong-Lin Chen et al.· 0 citations
Remote identity verification assumes that a camera stream shows a live person, yet public real-time face-swap tools can replace that person before the stream reaches the verifier. Passive deepfake detectors generalize poorly to new generators, while interactive challenges are often selected without evidence of which ge...
Anton Firc, David Drtil, Martin Perešíni et al.· 1 citation
We propose a morphing-based pseudo camera-on representation that transfers a user’s facial movements to a photorealistic morphed face, thereby communicating reactions without directly displaying the user’s unaltered appearance. To examine an appropriate degree of personalization, we evaluated five levels of resemblance...
Face image quality assessment (FIQA) estimates the utility of a face image for automated face recognition and has been proposed as a morphing attack detection (MAD) signal that requires no morph-specific training. We evaluate 17 FIQA methods from all current FIQA categories on 20 morphing attack sets from two datasets,...
Hao Peng, Bo-Chen Fu, Mahmoud Elshirbeny et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.