Gurudev Lipsync Shootout

How to evaluate this comparison

Goal: pick the lipsync model we'll integrate into the gurudev-knowledge German dubbing pipeline. The same 23-second source clip + same VibeVoice German audio went into every model — only the mouth synthesis differs.

1. Watch in this order

1. Original (English) first — anchor your perception of Gurudev's natural mouth shapes, head movement, and lighting.
2. Listen to the German audio alone (top of page) — get the pacing in your ear before judging mouth sync.
3. LatentSync 512² → MuseTalk → VideoReTalking → Wav2Lip → LatentSync 256². Top to bottom in the grid is roughly highest-quality to fastest.
4. Re-watch each one twice: once for sync (close your eyes for the audio, then look at the mouth on second viewing), once for naturalness (does the face look "right" or uncanny?).

2. What to look for

Criterion	What to check
Lip sync accuracy	Do mouth shapes hit consonants on time? Watch "besser", "gierig", "unglücklich" — the plosives are the giveaway.
Visual sharpness	Is the mouth crisp or does it look smeared/blurry compared to the rest of the face?
Identity preservation	Does it still look like Gurudev? Watch the beard line, chin, nose-mouth distance.
Temporal stability	Does the mouth jitter or "boil" between frames? Pause and step frame-by-frame if unsure.
Edge artifacts	Look at the boundary where the synthesized mouth meets the rest of the face — any visible seams?
Naturalness	Gut feeling — does the result feel like a real person speaking, or does it cross into uncanny valley?

3. Quick scoring

For each model, give it a 1–5 on each criterion and add notes. The winner is the one I'll wire into the gurudev-knowledge pipeline. If two are close, we'll let speed and reliability break the tie.

4. Tips for viewing

• Use full-screen (click the expand icon on the video player) — the mouth detail matters and is hard to judge in the small grid view.
• Loop short sections by dragging the playhead — sync issues are most visible in the first 5 seconds where Gurudev says "Aber einer ist besser als der andere".
• Use 0.5× playback speed (right-click → speed) to see frame-level mouth shapes.
• Compare on a desktop/laptop, not a phone — the source clip is only 320×240 upscaled, so artifacts are subtle.

5. After you decide

Tell me the winner and I'll: (a) integrate it into the gurudev-knowledge pipeline as a new _de_lipsync.mp4 output slot, (b) batch-process Bhakti Sutras 1–20 over the next few nights, (c) re-publish here for QC. If none are good enough, we still have options: MuseTalk with the original dwpose path via WSL Docker, or a higher-res source clip (current source is only 320×240 upscaled — a 1080p original would let LatentSync 512² shine more).