Tested 11 October 2026 on an RTX 4090. On this matched sample, MuseTalk 1.0 took 27.834 s and 1.5 took 28.118 s: 1.5’s median elapsed time was 1.0% longer. Four measured runs per version; both output videos passed validation.
Measured comparison
| Measurement | MuseTalk 1.0 | MuseTalk 1.5 |
|---|---|---|
| Elapsed time: median (range) | 27.834 s (27.459–28.558 s) | 28.118 s (27.361–28.495 s) |
| Peak PyTorch allocated VRAM: range | 6.899 GiB | 6.899 GiB |
| Peak sampled process VRAM: range | 7.467–7.475 GiB | 7.463–7.475 GiB |
| Video resolution / frame rate | 512 × 512 / 25 fps | 512 × 512 / 25 fps |
| Video frames / container duration | 90 / 3.632 s | 90 / 3.632 s |
| Successful measured runs | 4 / 4 | 4 / 4 |
| Run failures | 0 | 0 |
| Output validation | Video + audio; full decode PASS | Video + audio; full decode PASS |
RTX 4090, 24,564 MiB VRAM; Ubuntu 24.04; driver 590.48.01; Python 3.10.18; PyTorch 2.0.1+cu118; MMCV 2.0.1. Both used FP16, batch 2 and four CPU threads. Same stationary fictional portrait video and synthetic speech, with identical file hashes. One warm-up per version was excluded; measured order alternated across four pairs.
The timing ranges overlap. Elapsed time covers fresh-process startup, model loading, preprocessing, generation and encoding, with warm file caches and identical timing instrumentation. Install, downloads and transfers are excluded. VRAM peaks include model loading before FP16 conversion. The pre-existing ComfyUI process remained resident; its memory is excluded from sampled MuseTalk process VRAM. Memory figures describe this test, not a minimum GPU requirement.
Output videos
The VAE samples latents, so outputs vary between runs. Neither version used a fixed seed. Both displayed videos use measured run 1.
Same-frame screenshots
Frames 20, 40 and 60, selected before testing: 0.8, 1.6 and 2.4 seconds at 25 fps. Full frames retain the output pixels. Use the videos to assess motion and audio alignment; no numerical lip-sync or identity-quality score was measured.
0.8 seconds — frame 20
1.6 seconds — frame 40
2.4 seconds — frame 60
Model and setup differences
| Difference | MuseTalk 1.0 | MuseTalk 1.5 |
|---|---|---|
| Training | Original public L1-loss version | Perceptual, GAN and sync losses; two stages; spatio-temporal sampling |
| Upstream quality claim | Baseline | Improved clarity, identity consistency and lip-speech synchronization; not quantified by this test |
| Processed face region | 256 × 256 | 256 × 256 |
| Checkpoint | models/musetalk/pytorch_model.bin |
models/musetalkV15/unet.pth |
| Checkpoint download size | 3,400,076,549 bytes | 3,400,074,924 bytes |
| Config path | models/musetalk/musetalk.json |
models/musetalkV15/musetalk.json |
| Version argument | --version v1 |
--version v15 |
| Bounding box adjustment | bbox_shift from config/argument; tested at 0 |
Fixed at 0 in this code revision |
| Lower crop margin | No additional v15 crop margin | 10 pixels by default; kept in this test |
| Blending | Default branch | Jaw parsing; cheek widths 90 by default |
Switch checkpoint, config and version together. Both versions share the environment and supporting weights. Cropping and blending follow their official version branches, so this compares complete pipelines rather than isolated checkpoints with identical preprocessing. This short stationary input does not establish results for moving faces.
Downloads and exact versions
Download evidence and reproduction kit · Download all run measurements. Includes permitted inputs, labelled outputs, logs, environment, model checksums, timing records and reproduction scripts.
| Version | Official checkpoint | Official config |
|---|---|---|
| 1.0 | pytorch_model.bin | musetalk.json |
| 1.5 | unet.pth | musetalk.json |
Tested 11 October 2026. Code commit: 0a89dec45a0192b824e3cf4daf96c239440c5ed8. Model revision: 2bcb936e2fddb4d86db4c62fd45b387d0c061571. Sources: official project, tested inference code, official model files. Setup: installation guide and supporting model weights.
1 thought on “MuseTalk 1.0 vs 1.5: Tested Comparison, Outputs & Downloads”