为什么值得关注

该项目展示了在移动设备上实现 AI 图像生成的实际可行性，经过精细的内存优化、详细的性能基准测试，并注重隐私保护，同时采用 MIT 许可。

适合谁使用

对设备端机器学习感兴趣的 iOS 开发者
注重隐私、希望本地生成图像的用户
拥有兼容 iPhone（15 Pro/16 系列或更新机型）的 AI 爱好者
学习 MLX Swift 进行移动端推理的开发者

典型使用场景

在无网络环境下离线生成图像，避免数据上传到云端
为个人项目原型化和测试设备端扩散模型
教育用途，探索移动设备上的内存调优和模型量化
创建自定义图像，如壁纸或概念艺术

项目优势

完全设备端推理——无云端、无账户，图像数据不离开手机
在 8 GB 设备上实测并验证了内存使用（SDXL-Turbo 峰值 ~4.25 GiB）
支持两种模型（SDXL-Turbo 和 SD-Turbo），可调节步数并支持中途取消
开源（MIT），提供通过 Xcode 和免费 Apple ID 安装的清晰说明

使用前须知

要求至少 8 GB RAM 的 iPhone（15 Pro、16 系列或更新）；旧机型被阻止运行
需手动从源码安装，无 App Store 版本；免费 Apple ID 证书每 7 天过期
首次启动需通过 Wi-Fi 下载数 GB 的模型权重，且应用必须保持前台运行以免中断

README 快速开始

Alis Studio Mobile

On-device, native image generation for iPhone — a SwiftUI app that runs text-to-image diffusion models entirely on the device with MLX Swift. No cloud, no accounts; your images never leave the phone.

The mobile sibling of Alis Studio (the Mac desktop app), sharing its design language — clay accent, cream surfaces, the pine mark, light + dark.

Generate	Gallery

Real images generated on an iPhone 16 Pro Max — SDXL-Turbo, 512², 4 steps.

Requirements

[!IMPORTANT] An 8 GB-class iPhone is required. The app runs a multi-gigabyte diffusion model fully in memory; on devices with less than 8 GB of RAM it shows an "unsupported device" screen and will not generate (the OS would otherwise kill it mid-generation). This is enforced at runtime.

	Minimum
Device RAM	8 GB — iPhone 15 Pro, 15 Pro Max, 16, 16 Plus, 16 Pro, 16 Pro Max, or newer. Not supported: iPhone 15 / 15 Plus (6 GB), SE, and all older models.
iOS	17.0 or later
Free storage	~8 GB for SDXL-Turbo (≈7 GB weights + ≈0.3 GB fp16-fix VAE), or ~2.5 GB for SD-Turbo. Downloaded once, then cached.
Network	Required on first run to download model weights from Hugging Face. Keep the screen on — the download is several GB and the app is killed if it's suspended mid-download.
Hardware	A physical iPhone. MLX has no Metal backend in the iOS Simulator, so generation only runs on a real device.

Models & on-device memory (validated on iPhone 16 Pro Max, 8 GB)

Model	Config	Peak resident	Latency (512²)	Quality
SDXL-Turbo	4-bit UNet · `sdxl-vae-fp16-fix` · 2–4 steps	~4.25 GiB	~10–13 s/img	highest (photorealistic)
SD-Turbo	4-bit UNet · fp16 VAE · 4 steps	~3.4 GiB	~6 s/img	good, fastest

Both stay well under the 8 GB device's jetsam limit at 512². The memory tuning came from on-device measurement: the VAE decode activation is the memory peak, so SDXL loads the fp16-stable VAE (the stock SDXL VAE NaNs in fp16); the UNet is 4-bit; and the MLX cache is cleared between generations to stop it accumulating.

Features

Prompt → 512² image, fully on-device (MLX Metal)
Model picker

项目描述

On-device image generation for iPhone (MLX Swift) — the mobile sibling of Alis Studio

alis-studio-mobile