MFLUX — MLX native implementations of state-of-the-art generative image models

TL;DR

MFLUX is a single-author, Karpathy-style “minimal and explicit” port of major open generative image models to Apple’s MLX framework so they run natively on Mac. As of v0.11.1 (Nov 2025) it covers Z-Image (6B), FLUX.2 (4B + 9B), FIBO (8B), SeedVR2 (3B + 7B), Qwen-Image (20B), Apple Depth Pro, and legacy FLUX.1 (12B), and ships features like multi-LoRA, image-to-image, ControlNet-Canny, fill / inpainting, depth conditioning, in-context editing, multi-image editing, virtual try-on, and SeedVR2 upscaling — all from a uv tool install CLI. Practical relevance for Luma: a single project that makes the “run the SOTA open image stack locally on a Mac with quantization + LoRA” baseline concrete, including the multi-image Qwen-Image-Edit-2509 path that’s been making the rounds.

Key claims

The project is a “line-by-line MLX port” of HuggingFace Diffusers / Transformers implementations, written from scratch in MLX with only the tokenizers borrowed [README §Philosophy].
Supported model families and sizes: Z-Image 6B (Nov 2025), FLUX.2 4B & 9B (Jan 2026), FIBO 8B (Oct 2025+), SeedVR2 3B & 7B (Jun 2025), Qwen-Image 20B (Aug 2025+), Apple Depth Pro (Oct 2024), legacy FLUX.1 12B (Aug 2024) [README §Models table].
Both distilled and base variants are exposed where they exist (Z-Image, FLUX.2, FIBO, FLUX.1), and LoRA finetuning is listed as a supported workflow — not just LoRA inference [README §Models table, §Features].
The CLI installs as a single tool (uv tool install --upgrade mflux) and exposes per-model entry points; a five-line Python API path is also documented via inline-script-deps uv run [README §Installation].
Feature surface includes: quantization + local model loading, multi-LoRA with scales and library lookup, metadata export/reuse, in-context editing, multi-image editing, virtual try-on, ControlNet (Canny), depth conditioning, fill/inpainting, Redux, SeedVR2 upscaling, FLUX ControlNet upscaling, depth-map extraction, FIBO prompt tooling with VLM inspire/refine [README §Features].
Five third-party projects already build on MFLUX: MindCraft Studio (macOS app), Mflux-ComfyUI, MFLUX-WEBUI, mflux-fasthtml, mflux-streamlit [README §Related projects].

Method

Two-layer architecture per model. The MLX layer reimplements the network forwards (DiT blocks, attention, VAEs, controlnet adapters, samplers) directly against mlx.core so weights run on Apple Silicon’s unified memory without the PyTorch → MPS shim. The HuggingFace layer is reused only for the tokenizers and for the upstream weight files (the README explicitly credits Diffusers and Transformers as the reference implementations being ported).

The CLI is per-model rather than one generic entrypoint — mflux-generate-z-image-turbo, etc. — which lets each model expose its own flag set (steps, quantization width, image inputs for editing, LoRA stacks). Quantization, LoRA loading, and multi-image conditioning are first-class CLI flags; the README pitches -q 8 (8-bit) as the default for the Z-Image puffin demo at 9 steps.

The v0.11.1 release that triggered this filing adds Qwen Image Edit with multi-image support — i.e. the local-MLX equivalent of the Qwen-Image-Edit-2509 multi-reference workflow that the wiki has tracked from the cloud side (see Qwen-Edit-2509-Upscale-LoRA — community photo restoration adapter for Qwen-Image-Edit-2509 and Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509 for community LoRAs targeting the same backbone).

Results

No benchmarks in the README. The artifact is a working tool, not a paper: relevant “results” are the model coverage table (7 model families spanning Aug 2024 → Jan 2026 SOTA, including FLUX.2 from Jan 2026 and Z-Image from Nov 2025), the breadth of features (multi-LoRA + LoRA finetuning + ControlNet + inpainting + multi-image editing + upscaling all in one CLI), and the adoption signal (5 third-party projects building on it, an active CI badge).

Why it’s interesting

This is the wiki’s first datapoint for a production-quality local MLX port of the full open image-gen stack, complementing the server-side serving-optimization picture in Run FLUX.1-dev three times faster: both are about “ship the open SOTA fast on the hardware you have,” but one targets a single Mac and the other targets autoscaled H100s. For Luma’s audience the interesting bits are (a) multi-image Qwen-Image-Edit-2509 now works locally on Apple Silicon, closing the loop on the cloud-only Qwen-Edit-2509-Upscale-LoRA — community photo restoration adapter for Qwen-Image-Edit-2509 and Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509 community LoRA experiments, and (b) Z-Image’s 6B single-stream DiT (Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer) is supported with the distilled “turbo” variant out of the box, giving a useful local baseline for the efficiency direction documented in Run FLUX.1-dev three times faster.