Video2Game Explained: How the Open-Source Video-to-Game Workflow Works

Written by

VideoToGame team

Updated 4 min read

What the open-source Video2Game skill is, how it turns an H3 Max gameplay video into a Three.js browser game, what it costs per game and where it falls short.

In this article
  1. The idea in one sentence
  2. How the workflow runs
  3. What it costs
  4. What you need to run it yourself
  5. Where it falls short
  6. How VideoToGame relates to it
  7. FAQ
    1. Is Video2Game made by MiniMax?
    2. Is Video2Game free?
    3. Can I play the example games?

Video2Game is an open-source skill that teaches AI coding agents to turn a short gameplay video into a playable browser game. It was published by Lovis Odin (@OdinLovis) on October 28, 2026, together with 31 example games you can play and the source on GitHub. It is the project that put the phrase "video to game" in front of a lot of people, so here is what it does and what it doesn't.

VideoToGame (this site) is not affiliated with the project. We describe it because it is the clearest public example of the approach.

The idea in one sentence

Coding agents are good at writing game logic but bad at designing how a game looks, while video models are good at the look but can't make anything playable. Video2Game lets a video model design the game first, then has the agent rebuild that video as code.

In the author's words on X, an agent can't design a very original little game on its own, so a video model generates the gameplay first and the agent turns it into a real game you can play in the browser.

How the workflow runs

According to the project README:

  1. Generate the gameplay video. Each of the 31 games started as a 10 to 15 second clip made with MiniMax H3 Max on fal. The repo includes the prompts and the rules that made those clips easy to rebuild.
  2. Study the video. A script splits the clip into frames, a contact sheet, a color palette and cut times. The agent writes a short spec from named frames: camera, hero, the one mechanic, the key moment and the look.
  3. Generate the assets from the video's own frames. Concept sheets, cutouts, image-to-3D models, rigging and animations, sound effects and music all come from models on fal, and every call is logged. Nothing is drawn by hand or downloaded.
  4. Write the game. Every game is a single index.html with Three.js and nothing else: no game engine and no other library.
  5. Match the look. Tone mapping, color grade, fog, bloom and grain are set from the first scene, and the agent compares the video and the game side by side at the same moments after each change.
  6. Play-test it. The agent plays its own game with a keyboard script, records the run at the real frame rate and picks the liveliest seconds. Falls, walls and dead stretches get fixed before anyone else plays.

It works with Claude Code, Codex, Cursor, Gemini CLI and other agents that support skills.

What it costs

The README gives exact numbers for the 31 example games:

fal calls1,013 (594 images, 194 3D models, 290 sounds and music tracks)
Total fal cost202 USD
Average per gameabout 6 USD
Range in the galleryfrom 0.79 USD (Paper Dragon) to 15.48 USD (Train Painter)

Those figures cover fal only. The coding agent (Claude, Codex or another model) is billed separately by its provider, and building a game takes many agent turns.

What you need to run it yourself

  • A coding agent that supports skills
  • Python 3.9 or newer, ffmpeg, and Node.js 18 or newer with playwright-core for recording
  • fal-client and a fal API key, which is billed per call

The install is a plugin command for Claude Code or npx skills add lovisdotio/video-to-game for other agents. Check the README for the current steps.

Where it falls short

  • Setup and keys. You need a local toolchain and paid API access before the first game.
  • Cost per game adds up. About 6 USD in fal calls on average, plus the agent, and some of the gallery games cost 10 USD or more.
  • Small games by design. The results are short, single-scene experiences built around one mechanic. That is the point of the method, but it is not a full game with menus, saves and many levels.
  • Results depend on the clip. The method works because the clips were written to be rebuildable: one camera, one hero, one mechanic. Arbitrary footage is much harder.

How VideoToGame relates to it

VideoToGame applies the same video-first idea as a hosted service: you describe a game or paste a clip, and the build runs on our side, so there is nothing to install and no API keys to manage. Each build waits in a queue, and its page shows each step as it runs. Read how to turn a gameplay video into a game for the general method, or start a build.

FAQ

Is Video2Game made by MiniMax?

No. It is an independent open-source project by Lovis Odin. It uses MiniMax's H3 Max video model through fal to generate the gameplay clips.

Is Video2Game free?

The code is open source, but running it is not free: the assets are generated with paid fal calls (about 6 USD per game on average in the README's gallery) and the coding agent has its own cost.

Can I play the example games?

Yes. All 31 are linked from the gallery site, and each one's source is in the repo's gallery folder.