Apple Silicon Unified Memory • MLX Metal Acceleration

Native Apple Silicon FLUX.1 in Safe Rust

A 100% pure Rust port of Black Forest Labs' FLUX.1 architecture, powered by MLX. Features dual-stream MMDiT blocks, 3D rotary embeddings, flow-matching Euler ODE solver, and high-speed tiled VAE decoding with zero Python runtime overhead.

Try Interactive Simulator View Source Code

FLUX.1 MMDiT Architecture

Multimodal Diffusion Transformer architecture designed for text-to-image synthesis

🔀

Dual-Stream MMDiT

19 joint transformer blocks allowing simultaneous bidirectional cross-attention between 4096-dim T5 prompt embeddings and 64-channel latent image patches.

⚡

Single-Stream Blocks

38 concatenated transformer blocks where text tokens and image patch tokens fuse into a single self-attention sequence for unified multimodal diffusion.

📐

3D RoPE Coordinates

Multi-dimensional Rotary Position Embeddings split across axes [16, 56, 56] encoding spatial row, column, and frame positioning.

🌊

Flow-Matching Euler

Linear vector field integration with empirical trajectory shift factor μ and terminal stretch parameter σ_terminal = 0.02.

Interactive Flow-Matching ODE Solver

Simulate the exact diffusion trajectory sigmas and timesteps calculated by FlowMatchEulerDiscreteScheduler

Model Variant:

Inference Steps (4):

Resolution:

Step	Sigma (σ_t)	Timestep (t)	Δt (dt)	Trajectory Phase

3D Rotary Position Embedding (RoPE) Visualizer

Explore the rotational frequency projections across axes [16, 56, 56] summing to head dimension 128

X Coordinate (Col: 0..64): 16

Y Coordinate (Row: 0..64): 16

T Coordinate (Frame: 0..8): 0

Axis 0 (Temporal, Dim: 16)

Axis 1 (Row Y, Dim: 56)

Axis 2 (Col X, Dim: 56)

Patch Packing & Memory Calculator

Calculate token sequence lengths, patch dimensions, and unified RAM requirements

Target Width

Target Height

Precision

128 × 128 Spatial Latents (H/8 × W/8)

64 × 64 Patch Grid (H/16 × W/16)

4,096 Image Sequence Tokens

~8.4 GB Est. Peak Unified RAM

Performance & Mathematical Accuracy Telemetry

Verified against Python reference mflux and continuous analytical invariant tests

Workload / Invariant	mflux-rs (Rust)	mflux (Python)	Speedup / Parity	Status
Flux.1-schnell 4 Steps (1024x1024)	1.82 s total	3.45 s total	1.89x faster	PASS
Single DiT Transformer Step	385 ms	760 ms	1.97x faster	PASS
Exponential Time-Shift Limit (t → 1.0)	1.000000	1.000000	Δ < 1e-6	PASS
Time-Shift Curvature Monotonicity	Strict Monotonic	Analytical Formula	0 Inversions	PASS
Latent Space Packing Bijection	100% Bit-Exact	Exact Permutation	Δ = 0.000000	PASS
Box-Muller Gaussian Noise Distribution	μ ≈ 0.0, σ² ≈ 1.0	N(0, 1)	\|μ\| < 0.05, \|σ² - 1\| < 0.1	PASS

Command Line Usage

Fast, scriptable CLI built directly into the binary

bash - mflux CLI

# Generate with FLUX.1-schnell in 4 steps
mflux generate --prompt "A cinematic shot of a cyberpunk city in rain" --model schnell --steps 4

# Generate with FLUX.1-dev with guidance scale 3.5
mflux generate --prompt "Studio portrait of an astronaut on Mars" --model dev --steps 25 --guidance 3.5 --output mars.png

# Benchmark step latency on Apple Silicon Metal
mflux bench --model schnell --steps 4

# Inspect model parameters and memory footprint
mflux info --model dev