Prueba de sincronización de labios con IA basada en audio

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. ¿Por qué Wav2Lip enmascara la mitad inferior de la cara objetivo?
   • Para reducir el tamaño del archivo
   • To hide the actor's identity
   • To force the model to generate the mouth from the audio instead of copying it
   • To speed up face detection
2. ¿Qué hizo que el discriminador de sincronización de Wav2Lip fuera diferente de enfoques anteriores?
   • It was a pre-trained, frozen SyncNet-based expert rather than trained jointly with the generator
   • It listened only to background music
   • It judged image color only
   • It was trained after the generator finished
3. ¿Cómo se debe leer la LSE-D?
   • Higher means better sync
   • It measures video resolution
   • It counts dropped frames
   • Lower means better sync, since it is a distance between audio and video embeddings
4. ¿Cuál es el papel del marco de referencia en Wav2Lip?
   • It sets the audio volume
   • It supplies the person's appearance, such as skin and teeth texture
   • It determines the mouth shape
   • It selects the output language
5. ¿Cuál era una debilidad visual conocida del Wav2Lip original?
   • Blurry, low-resolution mouth regions
   • Refusing to process English
   • Changing the speaker's eye color
   • Removing the background
6. Which forensic cue can expose a lip-synced fake?
   • The video has subtitles
   • The speaker blinks
   • The lips fail to fully close on 'b', 'p' and 'm' sounds
   • The audio is in stereo
7. Why do visemes make audio-to-mouth mapping harder?
   • Visemes only exist in English
   • Visemes are audio files
   • Each phoneme has several unrelated mouth shapes
   • Many different phonemes share the same visible mouth shape
8. Why do production pipelines smooth face bounding boxes over time?
   • To change the speaker's identity
   • To prevent jittery crops that make the pasted mouth shake
   • To compress the audio
   • To add motion blur on purpose