Description
All you have is one product photo and you want a 3D model you can spin around. Normally that means a modeler or a multi-view reconstruction pipeline, plus cleaning up the background first. Idol collapses that into one command.
Give it a photo, a render or an AI-generated image of an object. It removes the background, frames the object, then generates up to 1,048,576 Gaussians, including the back and sides the picture never showed, and writes a standard 3DGS .ply that any splat viewer can open. The unseen sides are completed by the model.
One image in, one splat out:Armchairs, lamps, figurines, grapes and castles all come back with a plausible back side.
Built-in cutout:BiRefNet removes the background, so no pre-made transparent PNG is needed.
Two input modes:Use synthetic for renders and generated images and real for photographs; the crop margin differs.
Tunable detail:From 32,768 to 1,048,576 Gaussians, with adjustable steps, guidance and seed for reproducible results.
Fast:About 7.2 GB of VRAM on a consumer NVIDIA GPU and as little as two seconds per object; the default optimized sampling is roughly 1.7x faster.
Open source:Apache-2.0, about 3 GB of weights downloaded on first run. You must accept the terms of the DINOv3 and FLUX.2 encoders on Hugging Face first.
Give it a photo, a render or an AI-generated image of an object. It removes the background, frames the object, then generates up to 1,048,576 Gaussians, including the back and sides the picture never showed, and writes a standard 3DGS .ply that any splat viewer can open. The unseen sides are completed by the model.
Features
One image in, one splat out:Armchairs, lamps, figurines, grapes and castles all come back with a plausible back side.
Built-in cutout:BiRefNet removes the background, so no pre-made transparent PNG is needed.
Two input modes:Use synthetic for renders and generated images and real for photographs; the crop margin differs.
Tunable detail:From 32,768 to 1,048,576 Gaussians, with adjustable steps, guidance and seed for reproducible results.
Fast:About 7.2 GB of VRAM on a consumer NVIDIA GPU and as little as two seconds per object; the default optimized sampling is roughly 1.7x faster.
Open source:Apache-2.0, about 3 GB of weights downloaded on first run. You must accept the terms of the DINOv3 and FLUX.2 encoders on Hugging Face first.
