Viggle's Meridian rewrites camera angles on existing video

Meridian is an open-source, geometry-guided video tool from Viggle AI that lets you take an existing video clip or single still photo and reshoot it from entirely new camera angles and retimed speeds. You can orbit around a subject, pull off smooth lateral tracking shots, or freeze the action mid-stride for a Matrix-style bullet time take.

What it can do

Feed in a continuous 24 fps clip or a still photo, and you get complete freedom to map out camera paths in 3D space.

For instance, you can take a basketball dunk, keep the initial run-up at normal speed, freeze the athlete mid-air while swinging the camera 35 degrees, and then let the slam finish.

Outputs render at 24 fps on a 768-class canvas (1344 x 768 for standard 16:9 footage). Shot lengths range from 73 to 243 frames, delivering crisp 3 to 10-second generated takes.

How it works under the hood

Instead of guessing new angles out of thin air, Meridian relies on explicit 3D geometry combined with diffusion inpainting.

First, a geometry model called VGGT-Omega calculates depth and camera poses across input frames, turning the shot into a colored 3D point cloud.

Next, the system renders a rough reference video along your custom camera path. Regions occluded in the original footage are left as blank gray patches, allowing real-time path previews in the browser before heavy rendering begins.

Finally, the unmodified MiniMax-H3 base model takes both the original footage and the geometric reference video to fill in blind spots and polish the final image.

Viggle packaged this as two 2.5 GiB LoRA adapters rather than a bloated full checkpoint: a teacher adapter that understands geometry, and a 4-step turbo adapter for fast generation.

How to run it

Running Meridian locally requires serious hardware.

The base MiniMax-H3 weights weigh in at roughly 61.7 GiB, and the official pipeline needs a high-VRAM CUDA GPU. Quantization and CPU offloading are not supported out of the box yet, though the team is actively encouraging community ports to consumer cards like the RTX 4090.

All code, LoRA weights, and sample clips are live on Hugging Face.

You can run jobs via the Python CLI, spin up a local web-based Studio prototype on port 8412 for interactive keyframing, or use the official ComfyUI nodes and JSON workflows. Keep in mind that VGGT-Omega carries a noncommercial license and must be installed separately.

Why it matters

For video editors and visual effects creators, Meridian functions like an all-digital reshoot rig. Cinematic camera moves and bullet-time freezes that once required complex multi-camera arrays can now be generated from a single camera pass.

By anchoring the diffusion model to an explicit point cloud, it sidesteps the hallucination and warping problems that plague purely prompt-driven video generators.

While the hefty 60-plus GiB footprint currently locks out casual creators, we think Viggle's architectural decision—pairing lightweight LoRAs with a frozen base model—is brilliant. Once the open-source community ships functional 4-bit quantizations, this could easily become a staple workflow for video remixing and virtual production.