Meridian: A new perspective on space and time

One event. Anywhere. Anytime.

By Viggle AI

14 September 2026.

48-second teaser

Meridian is a geometry-guided video model for authoring new observations of existing events. Given a video, choose a new camera path and the source moments to observe. Follow the action from another angle, linger on a gesture, or hold an instant while the camera keeps moving.

Choose where. Choose when.

Where: design the camera's position, viewing direction, and lens over a shot.
When: let the action advance, hold a source moment, or change its pace by retiming the input.

Bullet time is one possibility, not the whole idea. Camera motion and source time can be composed into different ways of watching the same event. A single image can also be the starting point for a moving view.

The compound-camera example shows its source and requested path. The ballet examples use still images.

A dunk. Source action, revisited from new angles. An edited sequence; dunk completion is source footage.	Play. Hold. Resume. Linger on the splash, then let it continue.
Compose a camera path. Widen, orbit, slide, approach, retreat—one uncut take.	One image. Another viewpoint. A camera move from a single ballet image.

More examples · robots, animation, dance, and sport

Robot manipulation. The task continues from a higher viewpoint.	An animated event. Hold the action; move the camera.
Another ballet photograph. One image, a moving view.	A rising view. The gesture unfolds.
Through the powder. Move with the action.

Method

Input video enters VGGT-Omega; estimated depth and source cameras give colored 3D points. User-specified target cameras reproject those points into a warped video. Both the time-aligned source video and warped video condition Meridian to generate the output video. Video samples are real; 3D points and cameras are schematic.

Matched source, warp, and output frames; the 3D points and cameras are schematic.

The method is simple: use geometry to show a video model where to look.

1. Reproject the source. VGGT-Omega estimates depth and source-camera poses. We build colored 3D points, select the source moments, and project those points through an authored camera path into a warped video.

2. Generate the new view. Meridian, built on MiniMax-H3, takes the source video and warped video, aligned to the same source moments, and generates the new shot. Geometry guides the view; the video model fills missing regions and refines appearance.

Preview before generation. Once geometry is available, fast point-cloud rendering makes the chosen path visible. Check the framing, viewing direction, and uncovered regions—and adjust the camera before running the video model. This inexpensive preview is a useful consequence of making camera control explicit.

Why this matters

The shift is from generating another scene to choosing another observation of the same event. This is the world-model perspective behind Meridian: connect what we see to where and when we observe it, grounded in supplied footage rather than unrestricted simulation.

Unseen regions are generated, not recovered. Geometry errors and large moves—including 360° orbits—can destabilize the view. Time edits revisit supplied frames, and separate takes need not form a consistent world.

Try Meridian

Get started with Meridian.

Meridian uses MiniMax-H3 with precomputed text embeddings, without loading a text encoder. We welcome community work on quantization and CPU offloading toward smaller GPUs, including the RTX 4090; those configurations are not yet supported or validated by the provided implementation.

The release also includes a very basic, vibe-coded Studio demo to illustrate how to use the model—not a production editor. It supports multi-key camera paths and real-time 3D preview, not real-time video generation.

Watch the 38-second Studio walkthrough

Authoring and geometric preview only, with some operations and waits omitted—not a final generated take.