speridlabs/idol-sdk

What it solves

Turning a single photograph or render into a complete 3D model traditionally requires multiple shots, scanning, or manual modeling. Idol solves that by generating a full 3D Gaussian splat from just one image, including the sides you never saw. It also removes the background and crops the object automatically, so you get a clean, ready-to-use 3D asset.

How it works

Idol is a diffusion-based pipeline. First, a background-removal model (BiRefNet) separates the object from its background. Then, the cropped image is encoded using DINOv3 and a FLUX VAE to produce conditioning signals. A diffusion transformer with 370M parameters denoises over 50 steps using classifier-free guidance to generate a set of 3D Gaussians. An octree decoder and a Gaussian decoder refine the representation into a standard 3DGS .ply file. The whole process is optimized to run on consumer GPUs, with faster sampling on by default and PyTorch compilation speeding up repeat runs.

Who it’s for

Artists, game developers, e-commerce sellers, researchers, or anyone who needs a quick 3D asset from a single image. You don’t need a multi-camera rig or manual reconstruction tools. The project provides both a command-line tool and a Python API, plus a Gradio demo for interactive use. It works with real photos or synthetic renders — just pick the right mode.

Highlights

• One image in, a complete 3D Gaussian splat out, with up to ~1 million Gaussians.
• Generates the hidden sides of the object, not just the visible view.
• Runs on consumer GPUs (7.2 GB peak VRAM) and can produce results in under 2 seconds (depending on hardware).
• Offers both synthetic and real-photo modes, plus a Gradio demo with 93 examples.
• Flexible output: standard .ply format viewable in any 3DGS viewer, plus camera info and preprocessed input image.
• Open-source with weights on Hugging Face, built on TripoSplat and using DINOv3, FLUX VAE, and BiRefNet.