画像から 3D モデルへの生成

Image-to-3D generation turns a single image of an object into a textured 3D model you can use in a game engine, AR viewer or modeling tool. The fast modern approach is feed-forward reconstruction: one pass through a trained network, taking seconds instead of the minutes or hours required by older per-object optimization methods. The key design came from the Large Reconstruction Model (LRM), a 2023 paper by researchers at Adobe and the Australian National University. LRM uses a pretrained image encoder (DINO) and a large transformer that maps image features into a triplane: three axis-aligned feature planes that together describe a 3D field. A small network decodes density and color at any point, which can be rendered like a NeRF during training and converted into a mesh with marching cubes afterwards. TripoSR, released in March 2024 by Stability AI and Tripo AI under an MIT license, built on the LRM design and reported reconstructions in under a second on a capable GPU. Stability's later Stable Fast 3D added UV-unwrapped textures and material estimates. Another family first uses a diffusion model to imagine new views of the object, an idea pioneered by Zero-1-to-3, and then reconstructs from those views, as InstantMesh does. The central limitation is that one photo does not contain the back of the object. Everything hidden is a learned guess, shaped by training data such as the large Objaverse collection of 3D assets. Common, symmetric objects come out plausibly; unusual designs, thin parts, transparent or shiny materials and fine text often fail. Scale is unknown from a single image, so dimensions are not trustworthy. A frequent misconception is that this equals 3D scanning or photogrammetry, which measure real geometry from many photos. Single-image models suit concepts, placeholders and prototypes, not parts that must fit.