← All experiments
World models

Talking avatars

Our own model-training work for talking avatars: turning a single image into an animated, speaking character.

Talking avatars · A single image, speaking

What we trained

We trained our own models for talking avatars using MiniMax H3 and AquinF03. This is part of our world-model research: learning how a character’s visual state evolves over time from an image.

The starting point is a single image. The intended experience is a character that can speak, giving an interactive world a visible presence.

Research direction

A talking character is a focused setting for studying continuity. Its appearance needs to remain recognizable as motion and speech unfold. That makes it relevant to the same questions of persistent state and responsive generation that motivate our environment experiments.

Questions guiding the work

  • Identity over time. How can the character remain visually consistent across generated frames?
  • Speech and motion. How should facial movement correspond to the spoken output?
  • Continuity. How can transitions feel coherent rather than like disconnected clips?
  • Interaction. How can a generated character fit into an ongoing exchange with a user?

These are research questions, rather than claims of benchmarked performance.

Connection to the interaction layer

Our platform brings together agents, world state, perception, and action. Talking avatars explore the visible character within that loop, alongside our work on generated environments.

A future persistent character would combine this presence with context and goals. The avatar-training experiment is one part of that broader direction.

Project notes

The project repository documents the model adapter, training setup, evaluation reports, and local UI for single-image talking-avatar generation.

Project links

For live demos with interactive media

Try live demos.