WorldCrafter Turns One Photograph Into a Space You Can Walk Through

The model has a name for every camera move, and none of them are adjectives

Open the instructions for WorldCrafter, the video model Tencent's ARC lab released on Monday, and you will not find prompt advice about slow pushes or dramatic low angles. You will find a table. Forward. Backward. Left. Right. Up. Down. Turn left thirty degrees. Look up fifteen.

You write your camera move as a short list of those, one per line:

forward1x2
yaw_left30x3
backward1

Two moves forward, three left turns, one step back. That produces six pieces of video, thirty-three frames each. The distances are distances. The degrees are degrees. Nothing in there is open to interpretation by a language model having an opinion about your intent.

There is a second thing in the instructions that matters more, and it takes a minute to notice. The model remembers. If you pan away from a bookshelf and come back, the bookshelf is the same bookshelf. That is the claim the paper is actually built on, and it is the one that separates this from a year of camera-control add-ons that could move a lens but could not hold a room together.

What you can make with it

Feed it one photograph, or just a written description if you have no image. Give it a path. It renders a continuous walk through that space, chunk by chunk, holding onto what it has drawn so far.

For an illustrator or a previz artist, the use is obvious and immediate: you have a single concept frame and you want to see the room. For a documentary editor, it is a way to put motion into a still. For anyone building a set in their head before anyone builds it in the world, it is a rough space you can move around in.

The free demo is the place to test whether any of that is true for your own work. Nothing to install.

The catch is the size, and the name hides it

Here is the thing nobody writing about this release has put in a headline.

Two versions exist: WorldCrafter-Base and WorldCrafter-Fast. Fast is the one everyone will reach for. Its page on Hugging Face reports 147 gigabytes of files.

That number is worth sitting with. Inside it are two full generating models at roughly 57 gigabytes apiece, a 22.7 gigabyte model whose only job is reading your prompt, and a 2.5 gigabyte scene encoder. That is about 139 of the 147. The last eight are spread across five smaller files: two add-on files, two partial copies of the generating model, and the decoder that turns the result back into pixels. Fast refers to how few steps it takes per frame. It says nothing about what you download or what your machine needs to hold.

And it renders at 384 by 640. Smaller than a phone screenshot.

So the honest summary of the local version is: a hundred and forty-seven gigabytes of files, on Linux, on an NVIDIA card, with a specific build of PyTorch, to make a small video. That is not a criticism of the research. It is a caution against the word Fast doing work it was never meant to do. Read it as "fewer steps," not "runs on your laptop."

One more wrinkle, easy to trip on: the Base version reuses parts that live inside the Fast download. If you want Base, you keep both folders. And resuming a walk you already started, picking up a rollout where it left off, works only on Base. Fast starts fresh every time.

How to actually write a move

This is the part that transfers, whatever you end up rendering with.

The camera starts at the origin, facing forward. Forward and back follow wherever it is currently pointed. Up and down follow the world, not the camera, so tilting your view up does not make "forward" start climbing. Turning happens in place.

A few pieces of syntax earn their keep:

• x3 after an action repeats it. yaw_left30x3 is three separate left turns, which is ninety degrees across three chunks rather than a single lurch.
• & combines a move and a turn into one chunk. forward2&right2&yaw_left45 drifts diagonally while swinging left, which is how you get an arc instead of a staircase.
• reverse3 retraces the last three chunks. That is a there-and-back in one line.
• # starts a comment, so you can label sections of a long path.

The limit to respect: keep the total distance travelled in a single chunk to five or less. Longer moves get split into repeated actions. The model was trained on chunks, and asking one chunk to cover too much ground is asking it to invent more than it can.

You can also skip the written form entirely and supply camera positions directly as a numbers file, one position per frame, thirty-three per chunk, in real metric distances. That is the route if you are matching a move from a 3D package. The instructions are explicit that you should not pre-normalise that file yourself; the code handles it.

The prompt guide is better than the model's marketing

Tencent published a writing guide alongside the examples, and it is unusually specific about a problem most people will hit on their first try.

For a scene with something moving through it, open with a shot-description sentence. Their own example begins "A third-person gameplay-like camera closely follows" and then names the thing being followed, in their case a robot vacuum crossing a modern interior. Then describe the subject, its movement, and how it touches the space around it. Keep nearby objects and background landmarks nameable so the model has anchors as the subject moves.

For a static scene, do the opposite. Describe the space as a space: layout, foreground, background, what things are made of, how they are lit, how they sit relative to each other. Add no movement at all, because the camera path is already supplying the motion.

And then the detail I keep thinking about. One of their bundled examples is a stone chamber full of painted figures. The prompt for it explicitly says the figures are static, painted sculptures, that all of them remain motionless, that the poses are rigid and the garment folds fixed. They had to write that. Otherwise the model sees people and animates them.

That is a real craft note. When your scene contains anything that looks like it should move but must not, you declare it. The absence of a description is not an instruction to hold still.

Length: one focused paragraph, roughly 80 to 120 words, stretching to 100 to 130 when you are describing both a subject and a space. And do not write camera directions into the prose. You already wrote the camera. Two sets of instructions fighting each other is how you get a mess.

Where it breaks

The interactive demo does not work yet. The one with keyboard controls, where you walk around a scene like a game. The README says it is currently being debugged. That is the feature most people will want, and it is the one that is not ready.

Linux and NVIDIA only. Python 3.11, a CUDA 12.8 build of PyTorch 2.10. No Mac path. No Windows path in the instructions. If you are on Apple silicon, the hosted demo is your only door.

The licence is a question mark. There is no Apache tag, no MIT tag, nothing standard. The repository carries a LICENSE.txt with terms of use and third-party attributions, which means you have to open it and read it before you plan paid work around this. I am not going to summarise a licence I would want a lawyer to read. I am going to tell you it is not one of the permissive ones you recognise at a glance, and that this matters more than the resolution does.

Compiling is off by default. Turning it on speeds up long runs and makes your first render start much slower. Worth knowing before you conclude the model is broken.

And 384 by 640 is 384 by 640. Nothing here upscales for you. Whatever you make is a previz asset or a reference, not a delivery.

What I think this is for

Not finished video. Not yet.

What it is for is the thing that sits before finished video. You have one frame, from a photograph or a concept painting or a still you pulled out of an edit, and you need to know what the space around it feels like when you move. Blocking a shot. Selling a location that does not exist. Checking whether the geometry in your head survives contact with a camera.

The part that will outlive this particular release is the interface. Somewhere in the last few months the people building these tools stopped asking us to describe camera moves and started asking us to specify them, in numbers, the way anyone who has touched a 3D package has always specified them. WorldCrafter writes that as a shot list. The re-camera add-ons going around this week write it as measured geometry from your own footage. Different routes, same destination: the adjective is no longer the control surface.

I would spend twenty minutes on the hosted demo before I spent a hundred and forty-seven gigabytes on the local version. Write a three-line path. Write a prompt that names what must not move. See whether the room holds together when you come back to it. That answer is yours to get, and it will not look the same for a previz artist as it does for someone cutting a documentary.

---

Sources: WorldCrafter-Fast weights · WorldCrafter repository · Camera and prompt guide · Paper, arXiv 2609.24984 · Project page · Free hosted demo

Medium metadata Title: WorldCrafter Turns One Photograph Into a Space You Can Walk Through Subtitle: Tencent's new video model takes a written shot list instead of camera adjectives, and remembers what it already drew. The word Fast in its name is doing some heavy lifting. Tags: AI Video, Generative AI, Filmmaking, Previsualization, Open Source