Meridian Re-Shoots Your Footage From a Camera That Was Never There

Viggle's new video tool builds a second angle out of a single take. The most useful part takes a quarter of a second, and the part that will stop you has nothing to do with your graphics card.

Point Meridian at a clip you already shot, tell it to orbit fifteen degrees to the left, and it hands you the same moment from a camera position that did not exist when you rolled. Freeze the action at frame 24 and keep circling. Or feed it one photograph and get a move out of a still.

That is the pitch, and the demo reel is good. But the number on the model page that actually changed how I think about this is not in the results table. It is 0.24 seconds.

That is how long it takes to see whether your shot is going to work, before you spend the other thirty-six seconds finding out.

The grey holes are the whole product

Here is what happens when you hand Meridian a clip.

First it reconstructs the scene. A separate model from Meta called VGGT-Omega looks at your frames, works out roughly how far away everything is, and figures out where the camera was for each frame. Your footage becomes a cloud of coloured dots floating in space.

Then it renders those dots from wherever you asked the camera to go. And here is the part that matters. Anywhere your original camera never saw, there are no dots. The back of the actor's head. The wall behind the chair. The dust on the far side of the bike. Those areas come out grey. Empty. Holes.

Only then does the video model run, taking your original clip and that holed-out render as a pair of references, and painting in what is missing.

So the geometry pass gives you a rough, ugly, hole-riddled version of your shot in a quarter of a second. You look at it. If the holes are small, the render will probably hold. If half the frame is grey, you are asking the model to invent half your shot, and it will, plausibly, and wrongly.

I have used a lot of generative video tools that give you a prompt box and a wait. This gives you a coverage map. That is a different kind of tool, and it is the one an editor can actually plan around, because the failure shows up before you commit.

Viggle put the comparison in their own numbers: 0.24 seconds for the rough warp, about 36 seconds for the finished 73-frame take. You can audition ten camera paths in the time it takes to render one.

What you can actually ask it for

The camera controls are real controls, not prompt vibes. Reading the inference guide, it reads like a camera department, which is a nice surprise:

• --yaw 15 orbits fifteen degrees, positive going left.
• --truck 0.15 slides sideways. --boom 0.15 raises the camera.
• --dolly 0.8 pushes in. On its own it behaves as a dolly zoom, changing the lens as it moves so the subject holds its size. Add --zoom 1 and you get a plain push-in instead.
• --sweep ramps the move across the shot rather than applying it flat. --ease puts a soft start and stop on that ramp.
• --freeze 24:49 plays the first 24 frames live, holds frame 24 for 49 output frames, then lets the action resume.

That last one is the bullet-time move, and Viggle are careful to say it is one combination rather than the point. You can hold a moment and circle it, follow slow-motion action from a new angle, or pick a fresh angle on a sped-up sequence. Time and camera are separate dials.

One honest thing in the docs I want to flag, because it is the kind of detail most launch coverage skips. Pure optical zoom, a --zoom 1.5 with no movement, can get ignored by the model entirely. The guide says so and tells you to prefer moves with parallax. That makes sense once you know the mechanism: a zoom reveals no new geometry, so there is nothing for the reconstruction to work with. The tool is good at moves that change your position and indifferent to moves that only change your lens.

The two licences that decide this for you

Now the part I would want someone to tell me before I started a 64 GiB download.

The code is Apache 2.0. Free, permissive, fine. The weights are not. They are a derivative of MiniMax H3 and carry the MiniMax H3 Community License, which grants use and distribution of the weights and their outputs only inside what it calls the Applicable Territory. That territory excludes the European Union, the United Kingdom, the Republic of Korea and the United States.

Read that again if you are sitting in any of those four places, because it covers what you make with it, not only the files.

Second, VGGT-Omega is not included in the download at all. You request access from Meta separately, under the FAIR Noncommercial Research License. Viggle's own installation page is blunt about what that means in practice: the licence restricts commercial use of the research materials and their outputs, and "here those results include the geometry used to make the reference render." The Apache licence on Meridian's code does not remove that. If you want to use this in paid work you need a differently licensed way to reconstruct the scene, and swapping the front end is not a flag you can pass. It is integration work.

So: a tool that solves a real problem for filmmakers, released with an unusually candid set of documents, that a filmmaker in London or Los Angeles cannot legally ship work from. I do not think that is Viggle being cynical. They inherited MiniMax's terms and they wrote them out plainly instead of burying them. But it is the actual answer to "can I use this," and it is not "buy a bigger card."

Put this into practice

There is a hosted version, and it is the only sensible first step.

1. Open the official Space. Viggle put a Meridian demo on Hugging Face on 15 September, and it is running now. Upload a clip, design a path, look at the geometry preview, generate. Judging by the assets it ships, it is limited to the short 73-frame take, which is about three seconds. That is enough to learn whether the idea works on your kind of footage.

2. Prepare a clip that will not fight you. This is where people will lose their first hour. The tool reads frames by index and always writes at 24 frames per second. It does not convert your frame rate and it does not detect cuts. Give it one continuous shot, exported at a constant 24 fps:

ffmpeg -i clip.mp4 -vf "setpts=PTS-STARTPTS,fps=24" \
  -c:v libx264 -crf 18 -pix_fmt yuv420p -c:a aac clip_24fps.mp4

Then check how many frames actually came out, because the container label lies more often than you would like:

ffprobe -v error -select_streams v:0 -count_frames \
  -show_entries stream=width,height,r_frame_rate,nb_read_frames \
  -of default=noprint_wrappers=1 clip_24fps.mp4

For a standard three-second take you need at least 73 decoded frames.

3. Start small. Fifteen degrees, with a ramp and a soft start:

python inference/sample.py --video clip_24fps.mp4 \
  --yaw 15 --sweep --ease --out out/orbit

4. Look at render.mp4 before you look at anything else. Every run writes a handful of files, and four of them are the ones you care about. out.mp4 is your finished take. render.mp4 is the holed geometry. source.mp4 is your input after cropping. grid.mp4 puts all three side by side. Watch the grid. If the subject looks bent or the depth wobbles in the geometry pass, more render passes will not fix it. You need a different source shot or a smaller move.

5. Only then get greedy. Pick an orbit centre on your subject with --pivot 0.5,0.5 --pivot-lock, add a hold with --freeze, stack a slide and a push into one continuous move. If you want to go local, budget a full day: Python 3.12, CUDA 12.8, PyTorch 2.9.1, a pinned build of the Diffusers library that you must not upgrade, and a package called PEFT that the requirements file forgets to list.

Where it breaks

Large moves are the main one, and Viggle name a number: around 40 degrees is flagged as a caution point in their own guide, and they are explicit that it is a caution rather than a threshold. Past that you are asking for surfaces the camera never recorded, and you will get them. They will look right. They may not be right. That distinction is going to matter a great deal the first time somebody uses this on documentary footage.

The memory requirement is not negotiable in the shipped version. About 88 GB for a three-second take, roughly 113 GB for ten seconds, measured on a B200 with the service already loaded. No compression option, no spilling to system memory, no splitting across two cards, and the installation page says so directly while inviting the community to build those things. There is no verified consumer-card configuration and Viggle decline to imply one.

Separately generated clips may not join smoothly, so this is a shot tool and not a sequence tool. Retiming changes which frames get used, and it cannot recover motion that was never captured, so slow motion out of low-frame-rate footage will stutter for ordinary reasons rather than clever ones. The included Studio is described by its own authors as a basic, vibe-coded demo with no authentication, which is honest and also a reason not to expose it to the internet.

And the reconstruction is doing more of the work than the branding suggests. When the geometry is wrong, Meridian can often cope. When the geometry is badly wrong, it cannot. Your input shot quality is the ceiling.

What I keep thinking about

The preview loop is the idea worth stealing, regardless of what happens to this particular model.

Almost every generative tool I use asks me to commit first and evaluate after. Meridian inverts that for one specific decision: it shows me the shape of what it does not know, cheaply, before I pay for the guess. A grey hole is an honest answer. It says "nothing here was ever photographed, whatever appears next is invention." Most tools give you the invention with no marker on it at all.

If that pattern spreads, the interesting question stops being how good the generation is and becomes how legible its uncertainty is. I would take a slightly worse model that draws me a coverage map over a better one that hands me a confident render with no indication of which parts it made up.

Whether any of us get to use this particular one is a separate question, and the answer is currently sitting in a licence file rather than a spec sheet. If you are outside those four territories and you run it on real footage, I would like to know where your grey holes showed up.

---

Medium metadata

Title: Meridian Re-Shoots Your Footage From a Camera That Was Never There

Subtitle: Viggle's new video tool builds a second angle out of a single take. The most useful part takes a quarter of a second, and the part that will stop you has nothing to do with your graphics card.

Tags: AI Video, Filmmaking, Video Editing, Generative AI, Open Source

Suggested kicker: A quarter-second preview shows you the holes in your coverage before you render. That is the real release.