vork: A Free 3D and VFX Renderer Your AI Assistant Drives, Built to Put CG Into Real Footage Instead of Faking It
The most interesting line in vork's README is eight words long: "Real footage + rendered 3D; no generative video."
It sits under an example shot of a woman asleep at a table, taken from a free Pexels clip. Into that plate (VFX shorthand for the real footage you add things to), vork places a small ceramic-and-bronze object, tracks it to the moving camera, keeps it correctly layered as the woman's hand moves past, and lights it to match the room. Every pixel of the woman is real. Every pixel of the object was rendered. Nothing was guessed by a video model.
That is an unusual thing for an AI tool to brag about in November 2026, when most AI video pitches promise the opposite. And the part that makes it AI at all is stranger still: vork has no editor window. You never touch it with a mouse. Your coding assistant writes the scene, renders it, looks at the picture and rewrites the scene until the shot works.
So is this the future of VFX work, or a research toy with a good tagline? After reading the code, the license and the limits page, my answer is somewhere specific in between, and it depends on what kind of artist you are.
What vork is, in plain terms
vork is a free 3D renderer from Voyager, the company behind a desktop app that calls itself "the open harness for creative work." The repository went public on November 5, and version 0.1.0 shipped on November 6 as a research preview.
The README describes it as "a headless 3D and VFX engine built for agents." Headless means there is no interface to click. An agent, here, is an AI assistant that can run programs on your computer, such as Claude Code or a similar coding tool. The loop, in vork's words: "An agent writes a scene, renders it, looks at the result and revises the code. There is no editor to drive."
A scene is a plain text file. The starter example describes a blue sphere on a grey grid floor, a rectangular light and a camera, in one short file you can read top to bottom. vork turns that file into PNG stills or MP4 movies on your graphics card.
The examples folder shows how far that goes: a flock of a million simulated birds with the camera skimming through it, a black hole that bends light (the README calls it an approximation, not exact physics), a water simulation, a kitchen with a moving camera, and the plate shot above, which is the one that matters for anyone who composites for a living.
Why "no generative video" is the point
Here is the problem every editor and compositor has hit with AI video this year. You generate a shot. The client wants the object a little to the left. You ask for that, and the model gives you a new shot: different light, different hands, a slightly different face. A small note becomes a full reroll, and you cannot promise a client that the next version keeps what they already approved.
vork sidesteps that by keeping the two halves of a VFX shot separate, the way a real pipeline does. The plate is the plate. It never changes. The CG object lives in a scene file with a position, a material and a light. Move it left and you change one number. The assistant can make that change for you, but the change is still a line of text you can read, compare with the last version, and undo.
That is the honest version of "AI for VFX." The AI does the typing, the fiddly setup and the trial renders. The footage stays true, and the decisions stay visible.
How the plate shot is built
The example lives in examples/vfx-artifact, and reading its scripts shows exactly how much work vork does and how much it hands to other tools.
The preparation script downloads the Pexels clip, checks it against a stored fingerprint, crops it to 1920 x 1080, keeps eight seconds and saves it as plate.mp4. The tracking script expects exactly 200 frames at 25 frames per second and writes the solved camera move to camera-solve.json, a frame-by-frame list of where the camera sits and which way it points. The comments are frank that the lens and scale are estimates.
Then two free AI models do the jobs that usually eat a compositor's afternoon. SAM 2.1, Meta's cut-out model, draws a matte (a black-and-white mask that says which pixels belong to her) of the woman and a separate one of her hand for every frame. Depth Anything V2 estimates how far each pixel is from the camera. vork renders the object with those mattes and that depth in hand, catches its shadow on stand-in geometry, combines everything at full resolution, adds a little lens scatter, and writes the movie.
So vork is the renderer and compositor in the middle. The tracking and matte models around it are ordinary open tools the example wires together, and they need an NVIDIA graphics card with Python, PyTorch and the models' own code installed. The README does not say they run on a Mac. That matters for the next section.
Put this into practice
If you want to know whether vork belongs in your work, give it one afternoon and one question. Here is the order I would go in.
1. Check your hardware first. vork needs a recent NVIDIA card with hardware ray tracing on Linux (Ubuntu 22.04 or newer) or Windows 10 or 11, or any Apple Silicon Mac running macOS 26. Intel Macs and older cards are out. Oddly, so are the big data-center cards: the README says the A100, H100, H200, B200 and B300 "lack the required hardware ray tracing," so a rented cloud GPU may not work.
2. Install and read the manual the way the agent will. The README's install script puts a vork command on your system. Run vork --version, then vork help. The help topics are the documentation your assistant reads, so skim them yourself. One matters more than the rest: vork help limits. The README says to read it "before committing to a shot."
3. Start with the README's own prompt. Open your coding assistant in an empty folder and give it this: "Read vork help. Make a small scene, render a preview and inspect the image." Then give notes the way you would to a junior artist. "Make the light warmer." "Drop the camera to eye level." "Turn the sphere into worn brass." Watch the scene file change after each note.
4. Look at the plate example before you try your own footage. Read the vfx-artifact README and run its steps if you have the NVIDIA setup for the matte and depth models. The scripts are a working template for your own shots: swap in your clip, your object, your camera solve.
5. Judge it on your own shot, not the demo. Pick a five-second plate you already know well, with a clear surface for an object to sit on and something that passes in front of it. That is where the matte and the track have to hold up.
The README's last word on that example is the right attitude for all of this: "Review all output frames and the hand passes in slow motion before delivering." It adds that a good score from the tracker is not the same as an accepted shot. Your eyes decide.
Where it breaks
vork is honest about its gaps, which is one reason I take the rest of the README seriously. The gaps are real, though.
No fog, no smoke, no volumes. "Fog and participating volumes are ignored with warnings in this release." Anything atmospheric is out, and the README promises no date for it.
Big or glassy scenes can fail. "Large scenes and dense transparency can fail, including on Mac." A glass shop window or a crowd of translucent jellyfish is a bad first project.
New effects compile slowly. The first render of a new custom program can take "about 3–4 minutes" on NVIDIA while it builds. Your assistant will wait through that, and so will you.
The speed claims need their footnote. vork's tables show it finishing stills 1.38 to 1.66 times faster than Blender's real-time Eevee renderer, and movies 1.14 to 2.32 times faster. The README adds right away that these are "not equal-quality speedups": Eevee drew 64 passes per frame with no cleanup filter, while vork drew eight and used a filter to clean up the grain. Its slower, more physically accurate mode ran 5.81 and 1.99 times slower than Blender's Cycles on the two tests it reports. Treat the numbers as "fast enough to iterate," not "faster than Blender."
The Blender bridge is broken right now. The README says the helper "currently fails during camera preparation on Blender 5.2.1," and the bridge carries its own GPL license.
You are steering through an assistant. If your assistant misreads a note, it edits the wrong thing. The scene file makes that easy to catch, but only if you read it. And every round trip spends your assistant's usage, which adds up on a long shot.
The license is generous now and may not stay that way. vork is "source-available," not open source. The preview license allows free use for any purpose, "including commercial work," but it covers this version only, and later versions may carry paid terms. You may not sell or rent vork, redistribute a modified copy, or run it as a hosted service for others (delivering rendered shots to a client is fine). Fine for client shots today. Risky as the foundation of a product.
Who should try it this week
If you are a compositor or motion designer who already reads a bit of code, vork is worth an afternoon now. The scene-as-text idea, the real-plate example and the readable revisions are a better model of AI help than any prompt box I have seen this year, and the commercial-use terms let you test it on a paid job.
If you live in Blender's interface and have never opened a terminal, wait. The bridge is broken, the setup for the matte models is real work, and nothing here is faster than the tools you already know.
Either way, steal the principle. Keep your plate real. Keep your CG in something you can edit. Let the AI do the typing, and keep the judgment for yourself. Then decide, shot by shot, how much of the typing you want to hand over.
- Sources: vork on GitHub · vork README · vork 0.1.0 release · VFX on real footage example · vork LICENSE · Voyager