404-GEN

404-GEN is a decentralized AI-powered 3D content generation platform operating as Subnet 17 (SN17) on the Bittensor network. The platform enables users to generate 3D models from text prompts, with the goal of democratizing 3D content creation for gaming, virtual worlds, AR/VR/XR experiences, and creative applications. Users without any coding or 3D design experience can describe what they want and receive a generated 3D asset in about a minute.

The platform leverages a range of open-source 3D generative technologies, including Gaussian Splatting, Neural Radiance Fields, 3D Diffusion Models, and Point-Cloud approaches. Users can generate 3D models through a web app at 404.xyz/app, a Blender add-on, and a Unity plugin. The Unity integration is notable: 404-GEN is the first decentralized 3D model generation tool to become an official Unity Verified Solution, available through the Unity Asset Store and accessible to Unity's 1.2 million monthly active users. Generated assets can be output as Gaussian splats and seamlessly converted into meshes for use in game engines and 3D production pipelines. The platform also maintains a catalog of previously generated models at 404.xyz/catalog.

The long-term vision is to power AI-native games and virtual worlds where all elements, including 3D assets, dialogue, and sound, are generated at runtime. This would allow creators to simply describe a game or experience and have it built in real time, leveraging the broader Bittensor ecosystem for different types of AI generation.

404-GEN operates through a competition-based system with miners and validators. Miners pull generation tasks from validators, produce 3D outputs in PLY format, and submit their results for evaluation. Tasks can be either text-based (text-to-3D) or image-based (image-to-3D). Prompts are generated from a curated set of categories derived from industry analysis and marketplace trends, with LLMs used to create diverse and descriptive generation prompts.

Validators evaluate submitted models using a multi-layered approach. Each output must pass industry-standard benchmarks covering speed of generation, prompt similarity, and artifact detection. Beyond these baseline checks, miners compete in direct duels where two miners generate models from the same prompt, and a Visual Language Model (VLM) judges which output is better. Every duel is logged to a public dashboard with explanations provided by the VLM, creating transparency and a continuous incentive for miners to improve their models. This system is inspired by chess Elo ranking, where miners must consistently outperform peers to maintain their position.