Idol: a new image-to-3d model that generates a Gaussian splat in 2 seconds

On October 28, 2026, the model speridlabs/idol was published on HuggingFace with the image-to-3d task. It turns a single image of an object into a 3D Gaussian splat: it removes the background, crops the object and generates up to 1,048,576 Gaussians, including sides that were not in the photo. The result is saved to a standard 3DGS `.ply` file. The model weighs 2.9 GB, the largest file being denoiser.safetensors at 1.4 GB. The weights are distributed under the Apache 2.0 license, the architecture is DiT. Running it requires 7.2 GB of VRAM, and generation takes less than 2 seconds. Before the first run you need to accept the terms of two gated models on HuggingFace: facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev (only the VAE of the latter is downloaded).

What it means

Our reference on AI tools for 3D already lists several generators with the image-to-3d task, and Idol differs from them in the type of output. Stable Fast 3D and Hunyuan3D 2.1 produce mesh models, whereas Idol generates a Gaussian splat — a different way of representing a 3D scene, more often found in rendering and viewing tasks rather than in classic modeling. In terms of VRAM requirements, Idol is closer to Stable Fast 3D: 7.2 GB versus 6 GB for the Stability AI model. Hunyuan3D 2.1 is noticeably more demanding — from 10 to 29 GB depending on the generation stage. At the same time, Idol is lighter in weights: 2.9 GB versus 3.7 GB for Stable Fast 3D and 13.9 GB for Hunyuan3D 2.1. Idol stands out on licensing: Apache 2.0 permits commercial use without restrictions on audience or revenue. Hi3D v1.0 allows commercial use, but at a scale of 200,000 MAU or more or revenue of 1 million USD per month a separate license is required. Hunyuan3D 2.1 is limited to 1 million MAU and does not apply in the EU, the UK and South Korea. Stable Fast 3D is free up to 1 million dollars in annual revenue. Idol is a tool for quickly getting a Gaussian splat from a single image, with moderate hardware requirements and a permissive license. It will not replace classic generators for those who need a mesh for further editing, but it can be useful for viewing and rendering an object from different angles.