Bytedance has introduced DreamActor-M1, a new AI system for generating videos with more exact control over facial expressions and body movement. The system is aimed at a problem that has become central to AI video: making a generated person or character move in a way that is both directed by the user and visually coherent.
The company describes DreamActor-M1 as a system built around hybrid guidance, meaning several control signals are used together rather than relying on a single input. In practice, that gives the model separate ways to guide a face, a head, and a body while still trying to keep the final video coordinated.
What DreamActor-M1 is designed to control
DreamActor-M1 focuses on three closely related parts of performance: expression, head motion, and body motion. That matters because a convincing AI avatar is not just a face that changes shape. It also needs gaze, head orientation, posture, and movement that fit together.
At the center of the architecture is a facial encoder. According to Bytedance researchers, this component can change facial expressions without tying those changes to a person's identity or head position. The researchers present that as a solution to a common limitation in earlier systems, where controlling one part of a performance could interfere with another.
The demos described in the source show expressions and audio from one video being applied to both an animated character and a real person. That points to one of the system's main promises: performance transfer across different visual targets, while preserving control over the face and movement.
How the system guides faces, heads, and bodies
DreamActor-M1's control stack is split across several mechanisms. For head movement, the system uses a 3D model. Colored spheres are used to direct gaze and head orientation, giving the model a structured way to understand where the subject should appear to look and how the head should turn.
For full-body motion, DreamActor-M1 uses a 3D skeleton system. It also includes an adaptive layer that adjusts movement for different body types. The goal is to make body animation look more natural when the source motion and the target body do not match perfectly.
These pieces are important because video generation often breaks down when a model has to coordinate multiple signals at once. A face can be expressive while the head feels wrong. A body can follow a pose while the identity or clothing becomes unstable. DreamActor-M1 is built to reduce those conflicts by assigning different parts of the performance to different guidance methods.
Training gives the model more viewpoint flexibility
During training, DreamActor-M1 learns from images captured at various angles. The researchers say this helps the system generate new viewpoints from even a single portrait. When information is missing, such as clothing details or pose information, the model fills in those details intelligently.
The training process has three stages. First, the model learns basic body and head movement. Then it adds more precise control over facial expressions. Finally, the system is optimized as a whole so the different controls work together more smoothly.
Bytedance says DreamActor-M1 was trained on 500 hours of video. The training data was split evenly between full-body and upper-body footage. That balance reflects the system's broader aim: it is not limited to lip-sync or face-only animation, but also tries to handle larger body movement.
Where Bytedance says it stands against other systems
According to the researchers, DreamActor-M1 performs better than similar systems in both visual quality and motion control precision. The comparison includes commercial products such as Runway Act-One.
That claim fits the system's technical emphasis. DreamActor-M1 is not described merely as a tool for generating a person on screen. It is described as a way to make that person follow more specific instructions across several layers of movement.
For creators, that distinction matters. A system that accepts a portrait or target character is useful only if the resulting video remains controllable. If facial expression, gaze, head position, and body movement cannot be steered reliably, the output may look impressive in short clips but become hard to direct.
The limits are still clear
DreamActor-M1 is not presented as a complete solution to AI video control. The system cannot handle dynamic camera movements, object interactions, or extreme differences in body proportions between the source and target. Complex scene transitions also remain difficult.
Those limitations define the current boundary of the system. It can guide human or character performance with more precision, but it is still constrained when the scene itself becomes more complex or when motion involves the environment around the subject.
Bytedance, which owns TikTok, is also working on several AI avatar animation projects at the same time. Earlier this year, the company launched OmniHuman-1, which is already available as a lip-sync tool on CapCut's Dreamina platform. Other projects mentioned alongside DreamActor-M1 include the Goku video AI series and InfiniteYou portrait generator.
Taken together, those projects show a broader push toward controllable AI avatars and portrait-based video generation. DreamActor-M1 adds another piece to that effort: a system built to make faces, heads, and bodies respond to more detailed direction while acknowledging that scene-level complexity remains a hard problem.