DreamGaussian

Published as an ICLR 2024 oral, DreamGaussian cut text-to-3D and image-to-3D generation from the hour-scale optimization of DreamFusion-style methods down to a couple of minutes. It optimizes a cloud of 3D Gaussian splats under score distillation, then converts the result into a textured mesh and sharpens it with a short refinement pass in UV space. It proved Gaussian splatting could work as a practical generative representation, and many later projects reused its two-stage recipe. The last commit landed on January 2, 2024, shortly after Stable Zero123 and ImageDream options were added, and issues and pull requests now go unanswered. It can still produce results on an NVIDIA GPU, but only if you compile its custom CUDA dependencies, a modified Gaussian rasterizer, simple-knn and nvdiffrast, against an older PyTorch and CUDA pairing close to the torch 2.1 and CUDA 12.1 setup it was tested with. The code is MIT, though the diffusion priors it calls carry their own terms. For new image-to-3D work, feed-forward generators such as TRELLIS.2 (MIT) or Hunyuan3D 2.1 return cleaner meshes in seconds without per-object optimization.