A single image contains no direct evidence about an object’s hidden sides, so single-view 3D reconstruction is necessarily a problem of plausible completion—not simply measurement. Idol makes that tradeoff explicit: it synthesizes unseen geometry and appearance in a Gaussian-splat representation, with output detail adjustable from 32,768 to 1,048,576 Gaussians.
Key Capabilities
- It handles photographs as well as renders and generated images, with separate cropping modes for real and synthetic inputs. That makes it useful across both captured and designed assets.
- The output is a standard 3DGS PLY, so it can be inspected in Gaussian-splat viewers and used in workflows that already support the format.
- The model card reports 7.2 GB peak VRAM and benchmarks quality at both the input view and seven novel views. On the reported H100 test, it takes 6.8 seconds at 50 steps and 262,144 Gaussians; speed varies substantially by GPU.
- Its reported comparison with TripoSplat, TRELLIS, and LGM shows stronger scores on several novel-view measures, while LGM is faster in the tested configuration and has the highest input-view PSNR. The results highlight that reconstruction quality depends on more than matching the original view.
Who It’s For—and the Tradeoffs
A good fit for creators and developers who need a compact, rotatable 3D asset from one object image and can accept inferred—not verified—details on occluded surfaces. Look elsewhere if you need faithful recovery of unseen geometry, multi-object or scene reconstruction, or a GPU-independent hosted service. Local use requires an NVIDIA CUDA GPU and access to gated image encoders; reported latency also ranges from seconds to tens of seconds depending on hardware.
