Product Memo: Infinite Canvas Multi-Level Framework

1. Overview

Infinite Canvas is an open-source, local-first AI framework that automatically configures, fine-tunes, and runs state-of-the-art image and video diffusion models on Apple Silicon devices. By dynamically detecting available hardware, selecting the right model presets, and offering distributed training, Infinite Canvas provides a one-click solution for advanced AI use cases—all without sending private data to the cloud.

Core Technical Innovation

At its heart, Infinite Canvas solves a deceptively complex challenge: How do we transform everyday Apple computers into enterprise-grade AI infrastructure? The solution required rethinking three fundamental aspects of AI deployment:

1. How to harness Apple Silicon's unique unified memory architecture (Hardware Optimization)
2. How to make multiple Macs work together as seamlessly as a cloud GPU cluster (Distributed Processing)
3. How to make this power accessible to non-technical creative teams (Automated Configuration)

This memo outlines the core technology stack, our immediate use cases, and the longer-term vision for expanding private, on-device AI across multiple verticals.

Infinite Canvas is built on top of Apple’s MLX framework, leveraging Metal Performance Shaders (MPS) for low-level GPU acceleration. Our choice of MLX is key to enabling distributed training and advanced generative AI on everyday Apple Silicon hardware. Below, we detail why we use MLX (vs. MPS or Core ML alone), how it underpins our CLI tools (Stable Diffusion 3.5 and Flux1-dev), and the specific advantages for local-first creative workflows.

flowchart TB
    subgraph User["User / Creative Team"]
        UI[Prompts & Video Assets]
    end

    subgraph CLI["Infinite Canvas Framework"]
        ICLI[CLI / macOS App]
        Config[MPI Configuration & Host Setup]
    end

    subgraph Node1["Node 1: Mac Mini M4"]
        MLX1[MLX Runtime]
        MPI1[MPI Process Rank 0]
        GPU1[Apple Silicon GPU/ANE]
        MEM1[Unified Memory]
    end

    subgraph Node2["Node 2: MacBook M2 Max"]
        MLX2[MLX Runtime]
        MPI2[MPI Process Rank 1]
        GPU2[Apple Silicon GPU/ANE]
        MEM2[Unified Memory]
    end

    subgraph Distribution["MLX Distribution"]
        AllReduce[All-Reduce Operations]
        DataSplit[Data Parallel Processing]
    end

    subgraph Output["Model Processing"]
        Sync[Gradient Averaging]
        Final[Model Output]
    end

    User --> CLI
    CLI --> Config
    Config --> Node1
    Config --> Node2
    
    MLX1 --> GPU1
    MLX1 --> MEM1
    MLX2 --> GPU2
    MLX2 --> MEM2

    MPI1 <--> |TCP Links| MPI2
    
    Node1 --> DataSplit
    Node2 --> DataSplit
    
    DataSplit --> AllReduce
    AllReduce --> Sync
    Sync --> Final

    classDef primary fill:#e1eaff,stroke:#9eb9ff,stroke-width:2px
    classDef secondary fill:#fff5e1,stroke:#ffd591,stroke-width:2px
    classDef hardware fill:#ffe1e1,stroke:#ff9191,stroke-width:2px
    classDef output fill:#e1ffe9,stroke:#91ffb6,stroke-width:2px
    classDef mpi fill:#f1e1ff,stroke:#d591ff,stroke-width:2px
    
    class User,CLI primary
    class Node1,Node2 secondary
    class GPU1,GPU2,MEM1,MEM2 hardware
    class Output output
    class MPI1,MPI2,AllReduce mpi

Infinite Canvas leverages Apple's MLX framework to distribute AI workloads across multiple Mac computers, effectively creating a private AI cluster.

The system uses MPI (Message Passing Interface) to coordinate between machines - for example, a Mac Mini M4 and MacBook M2 Max can work together by splitting the workload, where each device processes a portion of the data in parallel.

When processing large video or image generation tasks, each Mac handles its assigned portion, and the results are efficiently combined through a process called "all-reduce" that averages and synchronizes the results across all devices.

This distributed approach allows creative teams to harness the combined computational power of their existing Mac devices, eliminating the need for expensive cloud GPU resources while keeping all data and processing local.

Note: MLX reduces communication overhead in multi-node Apple setups, accelerating generative tasks that would otherwise saturate VRAM

Framework Comparison: MPS, MLX, Core ML (Apple’s Machine Learning Framework)

Our Two CLI Tools

Distributed Training CLI

Deep Dive

1. MLX on Apple Silicon

2. Distributed Computing: Making Macs Work Together