ByteDance's Seed team just dropped Seed3D 2.0, an image-to-3D model that can spit out a fully textured 3D asset from a single photo. The update focuses on improving geometric accuracy and material realism, and its API is now live on the company's Volcano Ark platform.
For geometry generation, it uses a coarse-to-fine two-stage strategy: first, a large-parameter DiT (Diffusion Transformer) establishes a rough topological structure, then it refines sharp edges and detailed surfaces. On the material side, a Mixture of Experts (MoE) architecture boosts high-resolution detail, while a visual language model (VLM) prior enhances the stability of material decomposition under unknown lighting conditions. The output is a full set of PBR (Physically Based Rendering) texture maps that can plug directly into a standard rendering pipeline.
In a blind evaluation, 60 reviewers with 3D modeling experience used about 200 test cases to compare Seed3D 2.0 against competitors like Hunyuan3D-2.5/3.1, Tripo 3.0, Rodin Gen2, HiTem v2.0, and its own predecessor, Seed3D 1.0. Preference rates for geometry generation ranged from 65.1% to 98.3%, while preference for the full textured 3D assets consistently exceeded 69%.
For downstream tasks, Seed3D 2.0 can break down a 3D asset into functional components, add joint information, and output it in URDF format, making it compatible with simulation engines like Isaac Sim for dynamic interaction scenarios like robotic grasping. At the scene level, it supports text, multi-view images, or video as input, allowing you to combine multiple assets to generate a complete scene. This kind of 3D generation is part of a broader trend where AI models are tackling more complex visual tasks, similar to how Moondream's Lens fine-tunes vision models for $20 and a few dozen images or how OpenAI's ChatGPT Images 2.0 adds 2K resolution and a 'thinking' mode for image creation.