3D Graphics

Three.js117.7k★

Three.js puts 3D in a web page: product configurators, data visualizations, portfolio sites, browser games. It wraps WebGL and WebGPU (the browser's low-level graphics APIs) in scenes, cameras, lights and materials you can actually work with. MIT licensed and free, with no paid tier, no account and no cloud service. There's nothing to host. Run npm install three, import it, and render; it all runs in the visitor's browser. The extras live in addons: GLTFLoader for the standard 3D model format, WebXR examples for VR and AR, and helpers for physics libraries like Rapier and Jolt (the physics engine itself is a separate install). Solo dev or enterprise, it costs zero. React shops usually add React Three Fiber on top. The catch: after 16 years it's still version 0.x, and releases break things. The official migration guide lists API changes for every release and tells you to upgrade old projects 10 releases at a time, because deprecation warnings only last 10 releases. Pin your version. The other cost is learning 3D itself: cameras, lighting and GPU performance are a different world from the DOM. Babylon.js is the alternative if you want a more engine-like, batteries-included API.

Lottie32.4k★

Designers export from After Effects, developers drop the JSON file into the app, and it just plays: vector-based, resolution-independent, tiny file sizes. MIT, originally built by Airbnb. The workflow is: designer creates animation in After Effects, exports via the Bodymovin plugin as a JSON file, developer loads it with lottie-web. That JSON is typically 10-50KB where an equivalent GIF would be megabytes. Works on web, iOS, Android, and React Native. Fully free and open source. No paid tier for the renderer. The After Effects plugin (Bodymovin) is also free. LottieFiles (a separate company) offers a paid platform for hosting, editing, and collaborating on animations, but you don't need it. Solo to large teams, the library costs nothing. LottieFiles' free tier gives you basic hosting and a library of pre-made animations. Their paid plans ($20-50/mo) add team collaboration, more storage, and advanced editing. The catch: you need After Effects (or a compatible tool) to create animations. That's a $22.99/mo Adobe subscription. Not every After Effects feature is supported; 3D layers, expressions, and some effects won't export cleanly. And debugging a broken animation means digging through JSON by hand.

blender21.7k★

Blender is a full 3D creation suite: modeling, sculpting, animation, rendering, compositing, motion tracking, and video editing in one package. It competes directly with Maya, Cinema 4D, and 3ds Max, tools that cost thousands per year. The fact that this is free and open source is remarkable. Used in film production, game development, and architectural visualization by professionals and studios. The Cycles renderer produces film-quality output. Geometry Nodes opened up procedural modeling. The community is massive and the plugin ecosystem is deep. Autodesk charges $1,875/year for Maya. Maxon charges $719/year for Cinema 4D. Blender charges nothing. The catch: the learning curve is steep and the UI, while dramatically improved in recent versions, still has its own logic that takes time to internalize. Some industry-specific workflows (like CAD integration) are weaker than commercial alternatives.

filament20.7k★

Filament is Google's real-time physically based rendering engine, the part of a 3D app that turns models, materials, and lights into the picture on screen. It runs on Android, iOS, Windows, Linux, macOS, and the web through WebGL2 and WebGPU, with backends for OpenGL, Metal, and Vulkan. Apache-licensed and completely free. Google uses it in its own products, which is about the strongest endorsement a graphics engine can get. This is a rendering engine, not a game engine. No physics, no audio, no scripting, no full editor. You work in C++, with Java bindings for Android and a JavaScript path for the web, plus a material compiler you run at build time. Prebuilt releases ship through Maven Central and CocoaPods, so you don't have to compile it yourself unless you want to. It's built to be small and fast, with a clear bias toward mobile where every millisecond and megabyte counts. If you're building a 3D viewer, an AR feature, or a custom app that needs PBR-quality visuals without dragging in Unity or Unreal, this is the lean option. Solo and small teams: a great fit if you can write C++ or use the web bindings. Larger teams shipping cross-platform graphics get one engine that covers every target. The catch is that rendering-only means you build everything around it. Scene management, input, asset pipeline, gameplay logic: all on you. Filament draws beautifully and does nothing else. If you want batteries included, reach for Godot or a full engine instead.

bgfx17.6k★

bgfx is a rendering library that lets you write graphics code once and run it everywhere. It sits between your game or app and the actual graphics API, Direct3D on Windows, Metal on Mac, Vulkan, OpenGL, WebGL in the browser, and handles the differences so you don't have to. The "bring your own engine" philosophy means it does rendering and nothing else: no scene graph, no physics, no asset pipeline forced on you. There's nothing to host or pay for. You pull it into your C or C++ project as a library and build. The ops reality is the build itself: bgfx is a serious C++ codebase with its own shader compiler and toolchain, so getting it wired into your project takes real setup, not a one-line install. Once it's in, it's rock solid and has shipped in commercial games. Solo developers and small studios who want cross-platform rendering without committing to a full engine like Unity or Unreal get exactly that here. It pairs well with your own engine code. This is not the pick if you want batteries-included game development, you'd reach for Godot or Unreal for that. The catch: bgfx gives you a renderer, not a game. You're responsible for everything around it, and that's a lot of work if you're not already comfortable in C++ graphics land. The payoff is total control and no licensing strings, BSD-2 licensed, free forever.

orillusion5.2k★

Orillusion is a WebGPU-first 3D engine. While Three.js runs on WebGL (the older standard), Orillusion is built from the ground up for WebGPU, which means better performance, compute shaders, and modern rendering techniques like clustered forward rendering and GPU-driven particle systems. It's fully free under MIT. TypeScript-native, component-based architecture, with a physics engine, animation system, and media extensions (video textures, chroma key). The engine is designed to feel familiar if you've used Unity or other component-based engines. The catch: WebGPU support is still limited across browsers. Chrome has it, but Safari and Firefox support is incomplete or behind flags. That means your audience needs a modern browser. At, the community is tiny compared to Three.js (111K). Documentation is sparse, examples are limited, and if you hit a bug, you're likely reading the source code. This is bleeding-edge territory: exciting if you're experimenting, risky if you're building a product that needs to work everywhere today.

DiligentEngine4.5k★

Diligent Engine is a cross-platform graphics abstraction layer. Write code once and run it against Direct3D 12, Vulkan, Metal, and WebGPU on Windows, macOS, Linux, iOS, Android, and the browser. Shaders stay in HLSL and the engine translates them to MSL, GLSL, or SPIR-V per backend. This is infrastructure for graphics programmers, not a game engine. If you are writing a CAD tool, simulation, scientific visualization, or any real-time 3D application that has to run everywhere, you either use this, bgfx, or sokol, or you write platform-specific code for each backend. Diligent is more featured than sokol and more modern than bgfx. Solo and small teams: use it freely under Apache 2.0. Metal support on macOS, iOS, and tvOS requires a commercial license from DiligentGraphics. Vulkan via MoltenVK stays free. Contact them for pricing on the Apple backends since it is not published. The catch: low-level graphics is hard. Diligent abstracts the API surface but you still need to understand descriptor sets, command buffers, and pipeline state. This does not replace Unity or Unreal. It is what you reach for when those are too opinionated for what you are building.

nvidia-macos-driver4.2k★

NullMoth's driver gets modern NVIDIA GeForce cards running Metal on macOS 15 Sequoia and 26 Tahoe, on Intel Macs and OpenCore PCs. It covers GTX 16 and RTX 20 through 50 cards by translating Apple's shaders for NVK, Mesa's open Vulkan driver, on top of NVIDIA's own open kernel modules. Free, source included. Installing it means switching off the locks macOS uses to keep unsigned code out of the kernel: SIP bits, AMFI checks and Secure Boot. The 1401 Mac app edits your OpenCore config, backs it up, and adds a "Remove NVIDIA driver" boot entry. Keep it. Releases land several times a day; 2.3.2 broke enough desktops that 2.3.3 rolled back. No paid tier, just a tip jar. The license is PolyForm Noncommercial: personal, research and nonprofit use only. Solo tinkerers with a spare OpenCore box: free. Small and larger teams: no, the license bars commercial use. The catch: an anonymous developer, an account under a month old, unsigned kernel code with no public build pipeline, and three rounds of security fixes in four days. Through 1.x any local app could read and write kernel memory. Then 2.3.7 fixed a video decoder bug reachable from a web page. Weekend machine only.

3d-model-anything2.8k★

3D Model Anything turns a room into an editable 3D scene you can walk through in a browser. An iPhone app captures it with LiDAR (the depth sensor on Pro iPhones and iPads) and photos. Your AI coding agent then rebuilds it as Three.js code (the JavaScript library behind most web 3D) using the included skill. Wes Bos built it; the code is MIT and free. Setup is easy now: the iPhone app went live free on the App Store on October 24. Scanning needs a LiDAR iPhone or iPad (the Pro models) on iOS 17+, and scans stay on the phone until you AirDrop them out. The rebuild happens on your computer, where the skill runs in your coding agent and the Three.js viewer runs on Node. Nothing is paywalled, but the modeling runs in your own agent, so that subscription is the real bill, plus a Pro iPhone if you don't own one. Solo tinkerers with a LiDAR device: free. Small teams and larger shops: treat it as a demo workflow, not a production pipeline. The catch: this is not one-click scan to model. You babysit an agent through measure, model and compare loops, room by room. For a finished scan file today, Polycam gets you there faster.

usd-brep2.0k★

USD BRep is NVIDIA's newly opened CAD geometry engine: it builds exact NURBS solids, not triangle meshes, and runs booleans, fillets, sweeps, offsets and tessellation on them. Solids go in and out as OpenUSD, the 3D scene format behind Omniverse. C++ and Python, Apache 2.0, free. pip install usd-brep gets a CPython 3.12 wheel for Windows or Linux (x86 and ARM) with OpenUSD bundled. No macOS wheel. Source builds need Git LFS, Visual Studio's 2019 toolset on Windows and NVIDIA's repo tooling. One trap: never install PyPI's usd-core 25.11 alongside it; NVIDIA says importing both crashes. Versus OpenCASCADE: OCCT is LGPL, far more mature, and ships STEP and IGES translators, which most CAD pipelines need. USD BRep's Python package has no STEP or IGES importer. Pick USD BRep if your pipeline already lives in OpenUSD or Omniverse; otherwise OCCT, or CadQuery on top of it. Free at every team size. The catch: it's a week-old 1.0 that accepts no outside code contributions. The first issues filed report kernel bugs: a boolean union silently dropping a body, a fillet that hangs, one that corrupts a different solid, results that change between runs. Maintainers are responding. Treat it as promising, not proven.

cell-architecture-studio1.8k★

Cell Architecture Studio is a 3D viewer for biological cell models. Seven specimen types (plant, animal, neuron, white blood, epithelial, bacterial, muscle), rendered with Three.js, with an AI tutor panel that surfaces learning prompts and tracks mastery as you click through organelles. MIT, runs in any browser, deploys to Vercel. Self-hosting is just static hosting. Clone the repo, npm install, npm run build, drop on whatever serves files. GLB models for the cells live in the public directory. The project ships its own Playwright-based visual verification that catches blank renders and major layout regressions. Biology teachers building lesson materials, science museums looking for an embeddable demo, or developers learning React + Three.js + Drei from a real codebase: useful. Production educational platforms with curriculum integration: this is a starting point, not a finished product. It's a prototype. The author labels the status that way in the README. Several specimens use procedural fallback geometry instead of real GLB models. The 'AI tutor' is a UI shell, not a connection to a live LLM. Treat it as a reference implementation worth forking, not a polished product.

TripoSplat1.4k★

TripoSplat turns a single photo into a 3D model you can spin around. Feed it one image and it generates a cloud of 3D Gaussian splats, no multi-angle capture or photogrammetry rig required. The remarkable part: both the code and the model weights are MIT licensed, commercial use included. That's rare in image-to-3D, where most models ship research-only licenses you can't build a business on. You need a GPU and a weights download from HuggingFace, but the dependency list is tiny. No diffusers, no transformers, roughly 2,000 lines across two files. There's a ComfyUI workflow and a Gradio demo if you'd rather not touch code, and inference is a single script. This replaces paid image-to-3D services like Meshy, Luma, and Kaedim for a lot of jobs, and the commercial-friendly license means you can actually ship what it makes. Solo and small teams with a GPU: run it yourself. No GPU on hand: TripoAI's own hosted Tripo Studio has a free tier and then paid plans. The catch: one photo can't see the back of an object, so the hidden sides are guessed, sometimes well, sometimes not. And you get splats, not clean editable meshes. Great for viewing and rendering, not for dropping straight into a modeling pipeline.

aval1.4k★

AVAL is a new format for interactive video on the web: prerendered motion that behaves like a state machine instead of a timeline. Frame-accurate transitions driven by app state, real alpha transparency, and sequential decoding in the browser. It ships as a working toolchain (compiler, CLI, and a web component player), MIT licensed, from the Pixel Point agency. Using it is client-side: install the player element and drop it in. The build side is where the requirements stack up: Node 22+, and playback depends on WebCodecs and WebGL support, with Safari fixed only recently. Vanilla JS today; the React wrapper is still on the roadmap. Worth watching if you build marketing pages or product demos with heavy motion. Rive is the closest comparison for state-driven animation; Lottie covers vector animation without the video-like fidelity. AVAL's pitch is photoreal prerendered motion that still reacts to the user. The catch: this is an early technical preview. The 1.0 engineering release just landed, the API is still stabilizing, and the whole thing rests on one agency's continued interest. Build a toy with it now; don't bet the redesign on it yet.

zeno1.4k★

Physics and VFX simulation normally means expensive proprietary software. Zeno is a node-based visual programming framework for building those simulations. Picture a Houdini-like node graph but open source, built for technical artists and simulation engineers. MPL-2.0 license, C++ core. You connect nodes to build simulation pipelines: meshing, physics solvers, rendering, data processing. Supports OpenVDB, CUDA GPU acceleration, and real-time preview. The node system is extensible. Write custom C++ nodes for your specific simulation needs. Zenus Technology offers commercial licensing and support, though exact pricing isn't publicly listed. The open source version under MPL-2.0 is free for use. You can modify it but must share changes to Zeno's own files. Solo VFX artists or researchers: the open source version gives you a capable simulation platform for $0. Small studios: free, add custom nodes for your pipeline. Larger studios: contact Zenus for enterprise support and potentially proprietary extensions. The catch: and velocity means a small community. If you hit a bug or need help, you're largely on your own. The node-based paradigm has a learning curve, and documentation is limited compared to commercial tools like Houdini. For production VFX, Houdini is the industry standard. Zeno is the open source alternative for teams that can't or won't pay SideFX's licensing fees ($2K-4.5K/year).

camera-to-blender1.4k★

Point a phone at an object, tap the shutter, and about a minute later it is sitting in your Blender scene as a textured mesh. camera-to-blender is the glue holding that together: a FastAPI relay server, a phone-shaped web app, and a Blender add-on listening on a WebSocket. MIT licensed, and the code costs nothing. The generation does not happen on your machine. Each capture fires two calls to Tripo3D's API, an untextured preview mesh so the viewport has something to show immediately, then a textured PBR model that replaces it. Setup means running the server, opening an ngrok tunnel (phone cameras require HTTPS), and pasting a wss:// URL into the add-on panel. Gemini handles background removal and is optional. Here is the real number. Tripo sells credits at 100 per dollar. The preview costs 20 and the textured model costs 30, so every object you scan runs about 50 cents. New accounts get 300 free credits that expire in two weeks, which is roughly six scans. The catch: a lovely demo wrapped around a metered API. Free ngrok URLs also rotate on every restart, so you reconnect Blender each session.