Test date: 12 October 2026. Fictional permitted portrait and synthetic speech. Original assets unchanged.
15 direct PNG runs; 1254x1254, 25fps, 90 frames, 3.632 seconds; fully decoded.
Official source commit 0a89dec45a0192b824e3cf4daf96c239440c5ed8. HF revision 2bcb936e2fddb4d86db4c62fd45b387d0c061571.
Seed 20261012, RTX4090, batch2, strict deterministic math attention wrapper. No tracked upstream source edits.
All runs emit a post-write cleanup error about save_dir_full; MP4 exists and full decode passed. Original matrix stopped on v1 repeat hash mismatch; remaining three controls ran separately under same settings. This repeat variability limits causal pixel comparisons for v1.
These are one photo and one 3.63-second audio clip, not a quality ranking. Wall times include model loading and are diagnostic, not throughput benchmarks.
Single photo repeats its base frame. No natural head motion or blinks generated; lower face synthesis/blending can change cheeks. Internal synthesis uses 256x256 face crops; original canvas resolution does not imply that detail.
Screenshots are actual browser captures of original photo, saved configuration/log, generated output and output comparison viewer. Arrows only are added to separate copies. Raw originals preserved.
Paths in config/commands point to the test machine; replace paths for your own installation. Model weights are not bundled. Install official MuseTalk and its model dependencies.
