Sperid Labs releases Idol, a single-image 3D Gaussian model
Original title: speridlabs/idol — new model trending #30 on Hugging Face
Idol emits up to 1,048,576 Gaussians as a .ply, using about 7.2 GB of VRAM and under two seconds.
Sperid Labs published Idol, a model that turns one photo, render or generated image of an object into a 3D Gaussian splat, including unseen sides. It removes the background, frames the object and writes a standard 3DGS .ply. Weights are about 3 GB; the full request is about 1.80B parameters and peaks at 7.2 GB of VRAM. Inference needs gated DINOv3 and FLUX.2 VAE encoders, Python 3.10+, PyTorch 2.5+ and an NVIDIA CUDA GPU.
Unconfirmed: The benchmark table is cut off, so full scores are not available in the text.
- Default output is 262,144 Gaussians; the maximum is 1,048,576.
- The card lists 50 sampling steps and a default seed of 42.
- Modes are synthetic and real, with different crop margins.
- Component sizes include a 370.6M DiT and an 840.6M DINOv3 encoder.
- A comparison with TripoSplat, TRELLIS and LGM is only partly shown.
- Products
- idol-sdk
- Organizations
- Sperid Labs
From the original article
5 min read at huggingface.coimage-to-3d · 221 downloads · 142 likes · license apache-2. 0
By speridlabs · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.