each::labs launches its own video character replacement model

each::labs launches Lockstep, a video recasting model capable of replacing multiple characters while preserving movements and camera angles.

A stunt performer runs across a parking lot toward the camera. In the generated version, a woman in a blazer reproduces his run exactly, but the setting has become a school hallway. The movements, shot duration, and framing are preserved. Only the character and the environment have changed. This transformation illustrates how Lockstep, the new video model presented by each::labs, works.

The technology relies on recasting, a method that consists of replacing characters in an existing video using reference photographs. Unlike traditional video generators, which must recreate a complete scene from a prompt, Lockstep uses a previously shot sequence as a reference for movements, lip sync, and camera work.

The model is available on the each::labs platform, within its generation interface and via an API. Announced pricing starts at $0.0475 per second of video, depending on the resolution. A single photograph is enough to replace an entire character

How Lockstep works relies on two elements: a source video and a photograph of each character to be introduced.

The user first imports a sequence containing the movements they wish to preserve. This can be a choreography, a commercial, a narrative scene, or a handheld video.

They then add an image of the new character. The system reconstructs their appearance in the video while striving to maintain the original performer's gestures and movements.

The transformation is not limited to the face. The model can modify hair, body shape, outfit, and other visual features derived from the reference.

According to each::labs, a single clear and properly lit photograph is enough to provide a character's visual identity. However, the result remains dependent on the quality of this image, the complexity of the movements, and elements that are partially obscured during the sequence.

The system can also process animals, mascots, and stylized characters, extending its use beyond replacing human actors. Preserving movements rather than regenerating them

Lockstep's main promise involves fidelity to the original movement.

A video generation from a prompt can reproduce a general intent, such as a person dancing or running, without precisely adhering to the gestures, footwork, or duration of a pre-existing sequence.

Lockstep follows a different logic: the source video defines the choreography and rhythm that the new character must reproduce.

The company claims to preserve steps, rotations, expressions, and gestures at the exact same moments as in the original recording. Camera movements—including pans, fast tracks, or the shakes of a handheld capture—are also preserved.

The model also supports lip sync. When a character speaks or sings in the original video, the new face must reproduce the corresponding mouth movements.

This approach allows a previously recorded performance to be reused, rather than asking the model to invent a new one.

However, it does not guarantee a perfect reproduction of every movement. The published demonstrations are curated by each::labs and are not accompanied by independent measurements of body fidelity or lip-sync accuracy. Multiple characters replaced in a single generation

Lockstep can modify several subjects during a single operation.

In its API documentation, each::labs provides for up to eight reference images. Characters are designated using tags like `@Image1` or `@Image2`, which the user associates with the people present in the video.

This method makes it possible, for example, to replace two dancers with a royal guard and a matador without having to run two separate processes.

The tags serve to specify which character should occupy which place in the scene. The model then seeks to maintain this correspondence throughout movements and perspective changes.

The user can also choose to keep the original setting entirely intact. In this case, only the specified characters are modified, while the lighting, framing, and environment are supposed to remain identical.

Conversely, an optional prompt allows the transformation to be expanded. It becomes possible to modify the background, mood, or location where the action takes place, all while reusing the actors' gestures.

This combination gives the model an intermediate role between character replacement, motion transfer, and instruction-guided video transformation. A commercial shot once, then adapted with multiple actors

The primary commercial use case presented by each::labs involves creating variations of existing productions.

A brand can shoot a commercial with an actor, then produce several versions featuring different characters, without reshooting the scene.

The choreography, product placement, lighting, and camera movements all remain based on the same shoot.

This approach could simplify the production of variations intended for different markets, audiences, or seasons. It can also be used to produce multiple versions of a social campaign from a single shot.

In film and visual effects, the model could be used to replace a performer in an already filmed stunt, experiment with different characters, or prepare temp versions of a scene.

For social applications, the implementation is more direct: a user provides a photograph and can place themselves into an existing dance routine or video clip.

These use cases remain subject to one important condition: having the necessary rights to the source video and the permissions to use the likeness of the people depicted. From $0.0475 to $0.18 per second

Lockstep's pricing varies based on the output resolution.

Resolution · Announced price per second · Estimated cost for 10 seconds 480p · $0.0475 · $0.475 720p · $0.09 · $0.90 1080p · $0.18 · $1.80

These amounts correspond to the pricing schedule published on the model's technical spec sheet.

However, each::labs notes that the actual amount billed depends on the total cost of running the workflow. The price displayed prior to generation is therefore an estimate, subject to change depending on the processing performed.

The