paragraph,note
1,Matched comparison evaluation notes — 15 October 2026
2,"Scope: one permitted fictional still-face-derived video and a synthetic English utterance, repeated in fresh processes. No real head movement. No claim about other identities, languages, profiles, difficult footage or long-form stability."
3,"Observed in run 1 at frames 20, 40 and 60 (0.80, 1.60 and 2.40 seconds): MuseTalk produces wider mouth openings and visible changes around the lower face. The skin around its mouth and chin appears smoother. LatentSync retains more visible local skin texture; its mouth is slightly open at 0.80 seconds and close to the input's closed-mouth shape at 1.60 and 2.40 seconds. These are visual observations, not phoneme correctness scores."
4,"Subjective assessment: for this stationary sample, MuseTalk's articulation is easier to see. LatentSync's smaller edit should not be mistaken for better lip synchronization, and more visible MuseTalk movement does not prove greater accuracy. No listening panel, blind preference test, validated SyncNet evaluation or independent perceptual score was performed. The paired full clips are provided for review."
5,"Matched conditions: original byte-identical 512x512 video and mono 16kHz audio; native model settings and pipelines retained. LatentSync 1.6 uses 512px internal face synthesis, 20 steps, guidance1.5, seed1247 and official DeepCache option. MuseTalk1.5 uses 256px internal face synthesis, FP16 batch2, bbox0, extra_margin10, parsing jaw, cheek widths90/90. Distinct Python/Torch/CUDA environments, detector/preprocessing, precision handling, encoders and temporal chunking remain material differences."
6,"Timing: full fresh-process wall time, not pure network throughput or interactive end-to-end latency. Downloads/setup/transfer and independent decode checks excluded. Model/file caches may be warm, even on the first measured run, because setup validation reads models. Run order latent1,muse1,muse2,latent2,latent3,muse3; no thermal stabilization or randomized benchmark design. Short sample magnifies startup and compilation effects. The timing sampler polls nvidia-smi every0.5sec; sampled device memory includes existing idle ComfyUI (about265MiB total baseline) and is not an exact model allocation or minimum VRAM requirement. The baseline utilization counter can contain the previous run's sliding-window activity; no overlapping test model processes were launched."
7,Output duration: LatentSync emits92 frames/3.68sec video; MuseTalk90frames/3.60sec video. Both carry the3.632sec audio track. This follows their own audio feature/chunk handling. Raw outputs are retained; frame count is not normalized for timing. Shared source has91frames/3.64sec. Compare the common0–3.60sec interval; the combined presentation clip trims to3.60sec and re-encodes with labels provided in the page caption. It is not a raw benchmark output.
8,"Setup exceptions: decord0.6.0 has an embedded cp36 wheel tag; pip/uv flag it underPython3.10, despite working imports and actual decode. Constraint check has no missing/conflicting packages. ONNX initially fell back toCPU when libnvrtc.so.12 was absent from loader search path; the isolated environment's installed NVIDIA lib paths fixed it, and both detector sessions reported CUDA first with actual sample detection before timing. No upstream inference or checkpoint changes.
"