Genuine question. not even complaining, just curious.
Is the way Ltx works just fundamentally not very capable of producing a rigid dynamic of how teeth and mouth movement should render.. I understand it is probably the most difficult aspect of the human body to simulate, but with how good some results are compared to how bad other results are, it just seems like this part of the model just randomly falls apart or just doesn't quite know what to do. For the most part, mouth and facial expression are spot on, but the teeth often end up anomalous. I know that higher resolution and frame rate can often greatly improve this, and also having the subject larger in the frame, but what about a dynamic depth based noise scheduler or a rigid jaw/tooth constraint that could be either baked into a lora or the model itself.. Is anything like that even possible?
Not really,
We actually have a state-of-the-art dubbing pipeline built on LTX 2.3 (LTX 2.5 support is coming soon) that showcases the model's capabilities in this area.
You're right that all the points you mentioned — resolution, frame rate, subject size, etc. — have a major effect on result quality. Try using the spatial upscaler as a post-generation step. It doesn't just upscale the video — as a side effect, it also fixes some of the artifacts you mentioned.