Blog ·

IDOL online: turn one photo into a 3D Gaussian splat, in your browser

Sperid Labs' IDOL turns a single picture of an object into a 3D Gaussian splat, including the sides the photo never showed. What it does, its Apache-2.0 license, how to try it online and how to run it locally.

IDOL from Sperid Labs turns one picture of an object into a 3D Gaussian splat: it removes the background, frames the object and generates hundreds of thousands of small coloured 3D "splats", including the sides the picture never showed. The weights are open (Apache-2.0), but setting it up means a CUDA GPU, two gated image encoders and a Python environment. On LiveWeights you upload a photo and spin the result in your browser.

IDOLSperid Labs · Apache-2.0 (DINOv3 encoder under the DINOv3 License)One picture of an object in, a 3D Gaussian splat out, including the sides the picture never showed.Try it online →

What IDOL does

  • Single image in, full object out: geometry and colour all around the object, guessed from one view.
  • Gaussian splats, not a mesh: splats keep soft detail like fur, foliage and glossy surfaces that meshes struggle with.
  • Small and fast: a 370M-parameter diffusion transformer plus an octree decoder, about 7 GB of GPU memory and a few seconds per object.
  • Works on photos, renders and AI images: pick "Real photo" or "Render or AI image" so the crop fits your picture.

License: can I use it commercially?

IDOL's code and weights are Apache-2.0. It reads images through Meta's DINOv3 encoder, released under the DINOv3 License, which also allows commercial use. The splats you generate on LiveWeights are yours.

How to try it online

  1. Open the IDOL page and sign in, then pick a small credit pack (no subscription).
  2. Upload a picture of a single object: a product shot, a toy, a plant, a piece of furniture, or an AI-generated image.
  3. Choose the picture type and the detail (65k, 131k or 262k splats).
  4. Press Generate, then drag to orbit the object in the 3D viewer. Download the .splat to open it in SuperSplat or any splat viewer.

Tips for better results

  • One object, fully in frame, on a plain background works best; the background is removed automatically.
  • A three-quarter view shows more of the shape than a straight front or side view.
  • Change the seed to get a different guess for the hidden back side.

Run it locally

The code is idol-sdk on GitHub and the weights are speridlabs/idol on Hugging Face (about 3 GB). You need an NVIDIA GPU with 8 GB or more and access to the gated DINOv3 encoder. The pipeline writes a standard 3DGS .ply. Want a textured mesh instead of splats? Try Hi3D.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.