I've been running a lot of talking-photo renders (dogs, a teddy bear, a gorilla) through an audio-driven lip-sync model for short social clips, and animal faces failed far more often than human ones.
What fixed it, in order of impact:
- Close-up framing, face filling most of the frame, muzzle pointed at the lens.
- Mouth slightly open in the still. A closed mouth gave an almost static muzzle that reads like an off-screen narrator.
- Body upright. A dog lying with its chin on its paws never synced.
- A prompt that explicitly says the mouth opens and closes with every word, and "closed mouth" in the negative prompt.
Wide shots and side profiles were the consistent failures. Has anyone found a model that handles profile views for animals? (I build TalkPix, so my testing is on our own pipeline; interested in what others see elsewhere.)
What fixed it, in order of impact:
- Close-up framing, face filling most of the frame, muzzle pointed at the lens.
- Mouth slightly open in the still. A closed mouth gave an almost static muzzle that reads like an off-screen narrator.
- Body upright. A dog lying with its chin on its paws never synced.
- A prompt that explicitly says the mouth opens and closes with every word, and "closed mouth" in the negative prompt.
Wide shots and side profiles were the consistent failures. Has anyone found a model that handles profile views for animals? (I build TalkPix, so my testing is on our own pipeline; interested in what others see elsewhere.)