A conversational avatar brings several systems together: a conversation model decides what to say, text-to-speech produces audio, and a character system presents the reply. Keeping those responsibilities separate makes the implementation easier to reason about—and makes performance claims easier to interpret.
This article uses Mugen Morph as an example of the facial-performance stage. It describes the workflow and its evaluation boundaries. The proposed DeepSeek-and-TTS chat page remains an application concept, rather than a completed integration or a tutorial with reproducible code.
Mugen Morph: Bringing a character’s lip movements and expressions to life with audio
Start with speech and a character
The source workflow accepts a character image and an audio clip, with an emotion selected for the performance. It generates lip movements and facial motion from that audio and applies them to the character. The available character workflows include images, Mesh-based 3D digital humans, and 3D Gaussian Splatting (3DGS) avatars.
The evaluation question goes beyond whether the mouth moves. I would examine alignment with the speech, transitions between expressions, and whether the intended emotion remains readable throughout the line. A familiar character can make those changes easier to compare, but the same checks should also be repeated with the actual production character.
From character images and audio to expressive facial animation
Separate conversation from facial performance
For the proposed chat page, the sequence would be: user question, conversational reply, TTS audio, then audio-driven facial performance. DeepSeek is one possible conversation provider. Mugen Morph handles the facial performance; it does not generate the conversation itself.
This separation matters when measuring responsiveness. A fast animation stage does not establish the response time of the complete conversation. Reply generation, speech generation, delivery, animation, and rendering all contribute to what the user experiences. Those are measurements for a future integration, not results demonstrated by this article.
Presenting AI responses through a 3DGS avatar
Check continuity over the intended duration
Short generated clips can introduce discontinuities when stitched together: appearance may shift, or facial performances may meet awkwardly at a cut. Mugen Morph uses an autoregressive architecture to process audio continuously into motion representations, which can then be decoded into video frames.
The workflow is not restricted to fixed 15- or 30-second segments. Its image-based video output continues to use the supplied character image as the visual reference. That offers a reason to test longer dialogue, while still checking the actual output for appearance and expression continuity.
Long-form output example: observing character and facial-performance continuity
Read the measurements at the correct boundary
Figures in this article come from our current internal tests and estimates. Under selected configurations, video-generation cost can be roughly one-seventh of some other video-generation workflows. This is not a normalized benchmark or a universal price ratio.
| System | Reference time to generate a 10-second video |
|---|---|
| Mugen Morph | About 42 seconds |
| Seedance | Typically about 5 minutes |
| H3 | About 5–20 minutes |
The source reports fewer than 0.9B parameters for the combined motion-generation and video-decoding networks. It reports H3 at approximately 33B; Seedance’s parameter count is not publicly disclosed. Parameter count alone does not establish output quality or application latency.
The motion-generation stage can process 30 seconds of audio in under 3 seconds in the reported internal reference. Video decoding is a separate stage; this figure is not an end-to-end video-generation time or complete chat response time. Hardware, resolution, service configuration, and pricing affect the results.
Choose the output for the project
Image-to-video output suits prepared character dialogue. 3DGS, Mesh-based characters, and ARKit facial-animation data support other character pipelines. A conversational application would still need its dialogue and speech components.
Mugen Morph in video, 3DGS, and ARKit workflows
For a next experiment, I would measure a full reply, inspect a longer performance, and assess the animation on the intended character. Which boundary would you instrument first: conversation-to-audio, audio-to-motion, or final presentation?
Editorial note: AI assistance was used to adapt the wording of this article.






Top comments (0)