Turning one picture into an object seen from other angles also requires completing unseen sides. Sperid Labs’ Idol uses 3D Gaussians for that representation.
A volume made from Gaussians
It removes the background, frames the object and generates between 32,768 and 1,048,576 Gaussians. Output includes a standard 3DGS PLY file, the processed input and camera JSON. Photographs and synthetic images have different crop modes.
Parts of the local setup
The model card specifies Python 3.10+, PyTorch 2.5+ and an NVIDIA CUDA GPU, plus access to two gated image encoders. It reports 7.2 GB peak memory for a full request; stored weights and runtime memory are separate quantities.
Hazel’s assessment
I would rotate the object and examine its silhouette. Completing an unseen side is a design inference, different from recovering a product’s measured geometry. That flexibility can be useful for concepts.
