Twisterminigen: native Krea 2 Turbo studio on Swift and MLX for Mac

Twisterminigen is a native macOS application built with Swift, SwiftUI, and MLX that runs Krea 2 Turbo locally on a Mac mini M4 with 32 GB unified memory. It has no Python inside, no cloud inference, and requires users to accept the Krea 2 license before use. The pipeline uses Qwen3-VL-4B as text encoder, a 12.9B DiT, and Qwen-Image VAE. A subtle bug in the text encoder gave wrong positional indices to five tokens, which caused 99.32% pixel change after correction, even though cosine similarity tests looked perfect. The app supports multiple themes (Dark, Light, Glass), Queue, Presets, LoRA import, Regional Prompts, and immutable recipes that store all generation parameters. The developer used GPT‑5.6 Sol (Codex) to assist with most code changes. Weights occupy 31.18 GB, but memory is managed by sequential loading: Qwen, then DiT, then VAE. Best Fidelity q8 uses ~4.1 GiB more RAM than Default mixed-4/8. Character Sheet and Regional prompts produce coherent results via shared inference. The app is released as source-only under MIT license; model weights are obtained separately.