To mark our first anniversary, we open-sourced the tools and workflows we built on our film and game productions — one every day for seven days, from Sunday, October 18 to Saturday, October 24. New projects have kept coming since.
Every project is on GitHub. Try the source as it is, improve it, and build it into your own work.
Music and sound effects, made on your own machine.
Gives Claude Code, Cursor and any other MCP client a sound studio that runs on your own Mac or Windows PC. Ask in plain language and Stable Audio 3 generates the music or sound effect locally, then loops it seamlessly, sets the loudness (LUFS) and exports it for your game engine or video editor. No cloud, no API key, and nothing to pay per generation.
Water carves the sand; the sand reshapes the water.
A reference implementation of real-time, two-way coupling between a GPU shallow-water solver (Kurganov–Petrova 2007) and an erodible sand heightfield. Each sand cell carries its own wetness and strength, so a wet, packed wall stands against the surf while a dry, loose pile slumps and washes away — all from unscripted physics. Sixteen Godot 4.7 compute shaders, about 3 ms of GPU time per frame on an RTX 3060.
Does your generated sound match the prompt?
A local QA tool for generative audio. Built on Google’s multilingual EmbeddingGemma 2, which places text, sound, images and video in one shared space, it checks whether a clip matches its prompt, ranks your takes, runs technical QC on loudness, clipping and loop seams, judges how a sound fits the picture, and searches your own library. Use it from the CLI, a browser workbench or an MCP server.
Noisy frames, hushed in an instant.
A WebGPU denoiser that lets path tracing run in real time in the browser. It turns a noisy frame of one sample per pixel into a clean, flicker-free image. A tiny neural network of just 812 parameters (1.6 KB) predicts how to filter each pixel rather than the pixel colours themselves, so it never paints in detail that is not there. Denoising takes about 2.2 ms per 1080p frame on an RTX 3060, more than twice as fast as the standard SVGF method, and a small path tracer for three.js WebGPURenderer comes with it.
Forgot the file name? Find the picture by describing it.
Local AI image and video search for artists and filmmakers. Type “torii gate in the snow” and it finds that picture by what is in it, not by file name or tags. Search in Japanese, English or Chinese, show it an example image, or combine the two to narrow things down. Videos are searched shot by shot and play from the matching cut, and real production folders work as they are: RAW, PSD, EXR / HDR and more. Google’s EmbeddingGemma 2 runs entirely on your Mac, so nothing is uploaded. Use it from a web app, the CLI or an MCP server.
Not just what you said, but what you meant to say.
Fully offline voice dictation for the Mac. Hold a key, speak, let go, and the text appears wherever your cursor is. It recognizes 30 languages and 22 Chinese dialects and is tuned for Japanese, English and Chinese, including all three mixed in one sentence, with no modes to switch. Recognition runs on Qwen3-ASR entirely on Apple Silicon, so your voice never leaves the machine. Turn on the optional Polish step and it adds punctuation, drops filler words like “um” and, when you correct yourself, keeps only the correction. A free menu bar app, ready to install from a DMG signed and notarized by Apple.
One sentence to your AI, and your 3D turns cute and comic.
A free Blender add-on that turns your 3D models into cute comic-style toon art. Shadows take a colour from the palette instead of grey, hearts and stars are scattered where the light hits, and hatching, rim light and outlines are layered on top for the look of a printed manga. Highlights come in six shapes (dots, hearts, stars, little flowers, sparkles and moons) and there are 13 palettes. It works on characters, props, logos and backgrounds alike. Just ask an AI agent such as Claude Code or Codex, and it installs and applies the add-on for you; then fine-tune it in words while you watch the preview: “fewer hearts”, “make the edges glow”.
Above the water, below it, and right at the line between.
A real-time water renderer for the browser. It aims for a photographic look above the water, under it, and in the half-submerged “over-under” shot: an FFT ocean, light absorbed and scattered in the water, caustics rippling across the seabed, light shafts, and Snell’s window when you look up from below. A waterline that follows every wave exactly sweeps across the lens, and after the camera surfaces, a water film and drops run down the glass. Built on three.js WebGPURenderer and hand-written WGSL, with no game engine or plugins, it takes about 7 ms per 1080p frame on a laptop RTX 4070 and runs in desktop Chrome.
With generative AI, the tools for making films change almost daily. What lasts, we believe, is not which tool you used but how you combined and finished the work — the way things are made.
One year in, we want to hand over what we have learned making our own films and games in a form anyone can use. OPEN SOURCE WEEK was the first step.
The tools and AI production pipelines we build for our own titles are open source.
Visit our GitHub ↗
