topic,record
Scope,One permitted fictional stationary face and one synthetic English utterance; three fresh processes per model. No general quality ranking.
Input permission,The shared inputs were created for these tests/tutorials and may be reused; see the retained prior input permission record.
Wav2Lip compatibility failure,The official-folder non-GAN file is TorchScript. Stock training-checkpoint loader raised NotImplementedError before measured runs.
Resolution,Extracted 352 state tensors unchanged into the expected state_dict wrapper. Strict model load passed; CPU numerical comparison passed. Exact conversion and validation saved.
Environment adaptation,Original Python3.6/Torch1.1/NumPy1.17.1/Librosa0.7 pins were not reproduced. Isolated Python3.10.18/Torch2.0.1+cu118/NumPy1.23.5/Librosa0.9.2/Numba0.56.4/OpenCV4.9 used for Wav2Lip.
Timing boundary,"Fresh Python launch through exit, including imports/load/preprocessing/generation/encode. Provisioning, conversion, transfer, independent decoding excluded. Warm caches after setup; no cold boot claim."
Memory boundary,Whole GPU device used memory sampled at 0.5-second intervals; includes CUDA context and libraries; not allocator peak or minimum supported VRAM.
Fairness,"Same input bytes and hardware. Different native synthesis resolutions, FP32/FP16, detector paths, batch sizes, audio chunking and encoders. No compute-budget equalization or tuning."
Coverage limits,"Stationary head only; no head-motion, profile, occlusion, multiple faces, language coverage, live streaming or prepared-avatar measurements."
Setup effort,"Source/dependency/model preparation, one real checkpoint container mismatch and validated conversion. No comparative human setup-time measurement; no invented effort ranking."
Licensing,Upstream Wav2Lip restricts repository and results to personal/research/noncommercial use; current README is linked. No model weights redistributed.
Selection,"First successful run from each model; frames20/40/60 at0.8/1.6/2.4s, no cherry-picking. Identical160x144 crop at176,248. Presentation triptych trims3.52s common video interval; original files retained."
Subjective visual review,"At frames20/40/60, MuseTalk has crisper facial detail and wider mouth openings with visible teeth. Wav2Lip is softer over much of the face; its mouth stays closer to closed at20/40 and opens at60. Model and native encoder differences both remain; no isolated causal attribution or validated sync score."