sperid-swift-mlx ports Idol image-to-splat inference to Apple Silicon

Vincent Gourbin has released sperid-swift-mlx, a native Swift and MLX port of Speri Labs’ Idol model for Apple Silicon. From a single photograph of an object, the pipeline produces a 3D Gaussian splat entirely on the Mac, with no Python at run time. The public v0.3.0 release landed on 15 November 2026, followed the same day by v0.3.1, a dependency-range fix with no code change. The project is an independent community port and is not affiliated with Speri Labs.
The library runs BiRefNet background cut-out, DINOv3 ViT-H+/16 and FLUX.2 VAE encoders, a rectified-flow denoiser, and octree and Gaussian decoders. It writes a standard 3DGS PLY, a textured USDZ that passes usdchecker --arkit, and a GLB. On an M3 Max with 96 GB of memory, the author measured about 22 seconds per object with idol-turbo (five steps) and about eight minutes with the 50-step base model, at a process peak of about 11 GB. The weights are about 7.5 GB and are downloaded separately; Idol and Idol Turbo are Apache-2.0, while the gated DINOv3 and FLUX.2 VAE encoders keep their own licences.
The package includes IdolCore and a SwiftUI SplatView that renders the cloud with RealityKit, plus an idol command-line tool and idol-studio, a macOS app that mirrors the Idol Space. The code is MIT-licensed. The library and CLI require Apple Silicon and macOS 15 or later. The RealityKit splat viewer requires macOS 27, and building needs Xcode 27. A 16 GB Mac is untested.
«One Photo» Comic