Wav2Lip / MuseTalk reproducible research comparison — 16 October 2026

Same permitted fictional stationary face and synthetic speech as earlier MuseTalk tests. Input assets may be reused for these tests/tutorials. Model licensing remains separate. Wav2Lip upstream restricts its code and results to personal/research/noncommercial use; consult the linked current upstream licence statements. No model weights are included.

The archive retains six original output clips, exact commands/settings, timings, sampled GPU memory, source/model hashes, dependency freezes, setup failure and format-conversion validation, shared inputs and visual review notes. Public screenshot headings contain descriptive names, with no tracker/task IDs. Original browser captures and separate margin-arrow versions are retained.

Fresh reproduction requires an empty Linux /workspace, root in the GPU container, a working NVIDIA driver and supported Ampere CUDA GPU, ffmpeg, uv, tmux and network access. These scripts create isolated environments and install runtime OS libraries; they do not install a GPU driver. They are an exact record for the tested RTX3090, not a promise for every system. Keep at least25GiB free. Do not run setup on an existing production environment.

Place bundle/runtime and bundle/inputs in /workspace/d25-kit/runtime and /workspace/d25-kit/inputs. On a fresh machine run:
  bash /workspace/d25-kit/runtime/setup-all.sh
  python3 /workspace/d25-kit/runtime/run-matched.py

The revised setup includes the validated checkpoint container conversion. The first real setup attempt failed at the stock loader because the official non-GAN file was TorchScript. The unchanged 352 state tensors strictly match the upstream Wav2Lip model. Conversion/equivalence records preserve exact hashes; original inference/audio/model source trees were clean. The legacy requirements were adapted to explicitly retained modern compatibility environments.

Timers cover fresh Python process start to exit, including imports/load/preprocessing/generation/encoding. Setup and transfer are outside the timing boundary. Model files were read in setup and OS/library caches may be warm. These are not machine cold boots, prepared-avatar tests, live latency, or equal-compute comparisons. Video output lengths differ: Wav2Lip88frames/3.52s and MuseTalk90frames/3.60s; both container/audio tracks3.632s. Source91frames/3.64s. All original outputs preserved; the triptych trims to3.52s and is re-encoded only as a viewing aid.

One underlying short sample with three process repetitions per model. No validated automatic lip-sync/perceptual metric or blind study. Visual notes are subjective. Moving heads, profiles, other languages, long videos and streaming are untested. Sampled whole-device memory is not a minimum VRAM requirement.
