idol: Generate 1 Million 3D Gaussian Splats from a Single Image

Generates up to 1,048,576 3D Gaussian splats from a single image. It accepts rendered images, generated images, and real photos as input, automatically performing background removal and object framing. It outputs a standard 3DGS .ply file, including sides not visible in the photo.

1 / 8

Processing is possible in under 2 seconds using 7.2GB of VRAM on a consumer GPU. It utilizes a total of approximately 1.80B parameters across the DiT, Octree decoder, and Gaussian decoder, and applies torch.compile and cuDNN TF32 matmuls to increase sampling speed by approximately 1.7x. Request times were measured at 6.8 seconds on an H100 and 19.1 seconds on an RTX 4090.

In benchmarks against TripoSplat, TRELLIS, and LGM, it demonstrated competitiveness in PSNR, SSIM, and IoU metrics for the input view, as well as DINOv2 and CLIP scores for 7 novel views. In particular, the high DINOv2 (0.781) and CLIP (0.885) scores for novel views suggest excellent visual consistency from angles different from the input image. LGM was noted to have the limitation of producing blurry results despite high PSNR.

It runs on Python 3.10 or higher, PyTorch 2.5+, and an NVIDIA GPU environment with CUDA support. Access to DINOv3 and FLUX.2-dev VAE on Hugging Face is required, and approximately 3GB of weights are automatically downloaded on the first run. It is released under the Apache-2.0 license and is suitable for developers looking to add a single-image-based 3D conversion step to their 3D content generation pipelines.