NVIDIA

Edify 3D is a research project from NVIDIA’s Cosmos Lab that focuses on scalable, high-quality 3D asset generation. It is designed to address the growing demand for production-ready 3D content in industries such as video game design, extended reality (XR), film production, and physical simulation. Traditional 3D asset creation requires significant manual effort and specialized skills to achieve precise mesh structures, high-resolution textures, and accurate material maps. Edify 3D automates much of this process by generating detailed 3D models from text prompts or image inputs within approximately two minutes. The system employs a multi-stage pipeline: it first produces a coarse geometry, then refines it into a clean, topologically correct mesh with UV unwrapping, followed by high-resolution texture and material generation. Each stage is trained on large-scale datasets of 3D objects, enabling the model to generalize across diverse categories such as animals, vehicles, furniture, and architectural elements. The generated assets include albedo maps, normal maps, roughness maps, and metallic maps, making them directly usable in standard rendering engines like Unreal Engine, Unity, or Blender. Edify 3D also supports view-consistent generation, meaning the model ensures that textures and geometry remain coherent from all camera angles. This eliminates common artifacts like seam misalignment or blurry details seen in earlier methods. The framework uses a diffusion-based backbone combined with a transformer architecture to handle variable resolution inputs and outputs. It can generate assets with up to 4K texture resolution and polygon counts suitable for real-time or offline rendering. In addition to single-object generation, Edify 3D can produce asset variations, such as dif