Independent developer Vincent Gourbin has released sperid-swift-mlx, a Swift and MLX port of Speri Labs' Idol single-image-to-splat model for Apple Silicon.
Idol turns one photo of an object into a 3D gaussian splat, and Speri Labs distributes it with a Python SDK and a Hugging Face Space. sperid-swift-mlx v0.3.0 rebuilds the whole pipeline in Swift with no Python at run time, covering BiRefNet background cut-out, DINOv3 ViT-H+/16 and FLUX.2 VAE encoders, a rectified-flow denoiser, and octree and Gaussian decoders. Each stage was checked against the Python SDK on real weights, with octree anchors bit-identical and one denoiser forward pass within 7.7e-5 at fp32.
The denoiser runs either the 50-step base model or idol-turbo at 5 steps. On an M3 Max with 96 GB, the project measures idol-turbo at about 22 seconds per object and the base model at about 8 minutes, with a turbo process peak of about 11 GB. The weights are a 7.5 GB download and take about 4 GB in memory. A 16 GB Mac is untested.
sperid-swift-mlx writes a standard 3DGS ply, a textured USDZ for Quick Look and AR that passes usdchecker --arkit, and a GLB for Blender and the web. The idol export command meshes any existing 3DGS ply into USDZ or GLB. The IdolUI module displays the result as a RealityKit 27 entity through Apple's GaussianSplatComponent and a SwiftUI SplatView with orbit controls.
The package also ships the idol command line and idol-studio, a macOS app that mirrors the Idol Space with the same settings, the same 94 examples, a live 3D viewer, and an export menu. Both are built from source with Xcode, with no app store release. The library and command line require macOS 15 or later, the RealityKit viewer needs macOS 27, and the build needs Xcode 27. Gourbin's release notes say iPhone support comes later.
The code is MIT licensed. Weights download separately under the user's own Hugging Face account. Idol and Idol Turbo are Apache 2.0, DINOv3 falls under Meta's gated DINOv3 license, and the FLUX.2 VAE, taken from FLUX.2-dev, falls under the gated FLUX Non-Commercial License. The port is not affiliated with Speri Labs.
A v0.3.1 compatibility release followed the same day with no code change. sperid-swift-mlx is available now on GitHub.




