mflux

Flux image generation tool built natively on Apple's MLX framework

TLDR

Generate an image from a text prompt

$ mflux-generate --model [schnell] --prompt "[a sunset over mountains]"

Generate with the higher-quality dev model and more steps

$ mflux-generate --model [dev] --prompt "[prompt]" --steps [25] --seed [42]

Specify image dimensions

$ mflux-generate --model [schnell] --prompt "[prompt]" --width [1024] --height [768]

Save to a specific output path

$ mflux-generate --model [schnell] --prompt "[prompt]" --output [output.png]

Use a quantized model for lower memory usage

$ mflux-generate --model [schnell] --prompt "[prompt]" --quantize [8]

Generate using image-to-image with an init image

$ mflux-generate --model [dev] --prompt "[prompt]" --image-path [input.png] --image-strength [0.4]

Apply LoRA adapter weights

$ mflux-generate --model [dev] --prompt "[prompt]" --lora-paths [adapter.safetensors]

Save a local quantized copy of model weights

$ mflux-save --model [schnell] --quantize [8] --path [path/to/save]

SYNOPSIS

mflux-generate --model name --prompt text [options]mflux-save --model name --path dir [options]

mflux is a Flux image generation tool built natively on Apple's MLX framework, optimized for Apple Silicon (M1/M2/M3/M4). It generates images locally using Flux models without requiring a GPU server or cloud API. The package provides multiple CLI commands: mflux-generate for image generation, mflux-save for saving quantized model weights locally, and mflux-info for viewing image metadata.Installation is via pip (`pip install mflux`). Model weights are downloaded from HuggingFace on first use and cached locally. Custom models can also be loaded from local paths or HuggingFace repositories.schnell is faster with fewer steps needed (2-4 steps). dev produces higher quality but requires more steps (20-25). Quantization (4-bit or 8-bit) reduces memory usage for machines with limited unified memory. The --low-ram flag further reduces memory by releasing components after use.LoRA adapters allow fine-tuned styles and concepts to be applied on top of base models. Image-to-image generation is supported via --image-path and --image-strength.

PARAMETERS

--model, -m NAME

Model to use (schnell, dev, or a HuggingFace repo/local path).

--prompt TEXT

Text prompt for image generation. Use - to read from stdin.

--output FILE

Output image path.

--width PX

Image width in pixels.

--height PX

Image height in pixels.

--steps N

Number of inference steps.

--seed INT

Random seed for reproducibility.

--quantize, -q BITS

Quantization level (4 or 8 bit).

--guidance FLOAT

Guidance scale.

--negative-prompt TEXT

Text prompt for what the model should not generate.

--image-path FILE

Path to an initial image for image-to-image generation.

--image-strength FLOAT

How strongly the initial image influences output (default: 0.4, 0.0 = no influence).

--lora-paths FILE...

Paths to one or more LoRA adapter weights.

--lora-scales FLOAT...

Scales for each LoRA adapter.

--metadata

Export a JSON file with generation metadata alongside the image.

--low-ram

Reduce GPU memory usage by limiting MLX cache and releasing components after use.

--base-model NAME

Specify architecture (schnell, dev) when loading from a local path.

CAVEATS

Apple Silicon only (M-series Macs). Requires Python 3.10+. Model downloads are several GB. Memory usage depends on model and quantization level. The --low-ram flag helps on memory-constrained systems but limits generation to single images.

HISTORY

mflux was created by Filip Strand in 2024 to bring native Flux image generation to Apple Silicon using the MLX framework. It has since expanded to support multiple model architectures beyond Flux, including models loaded from HuggingFace repositories and local paths.