Make-It-3D

Why we included this project

Make-It-3D produces a full 360-degree textured 3D model from a single photograph, which makes it a useful reference for anyone in game art or product visualization who needs an asset when only one view exists. The two-stage pipeline first fits a neural radiance field under the frontal view, then converts the coarse result into textured point clouds, while a pretrained 2D diffusion model fills in plausible geometry and texture for the angles the camera never saw. The repo ships the training code for both stages and the test data behind the paper's results, so you can reproduce the outputs instead of trusting the teaser GIFs. Just plan for a fiddly environment: it wants CUDA-specific PyTorch, tiny-cuda-nn, and external tools like Stable Diffusion, DPT, SAM, and BLIP2. The codebase also works well as a base to extend, and since it draws on Stable-Dreamfusion, it is a handy comparison point in the diffusion-based 3D literature.