Mflux-ComfyUI (v2)

ComfyUI nodes for mflux 0.10.x (Apple Silicon/MLX)
한국어 | 中文

Overview

This fork upgrades the original nodes to use mflux >= 0.10.0 while keeping ComfyUI workflow compatibility (no graph rewiring required). It adds third‑party model support, richer quantization options, better LoRA handling, and a small MLX version hint in the UI.

• Backend: mflux 0.10.x only (legacy 0.4.1 runtime not supported)
• Graph compatibility: legacy inputs are migrated internally so your old graphs still work
• OS/Accel: macOS + MLX (Apple Silicon). MLX >= 0.27.0 recommended

Key features

• Quick text2img and img2img in one node (MFlux/Air → QuickMfluxNode)
• LoRA pipeline with validation (quantize must be 8 when applying LoRAs)
• ControlNet Canny preview and best‑effort conditioning
• Third‑party Hugging Face repo IDs (e.g., filipstrand/..., akx/...) with base_model selection
• Quantization choices: None, 3, 4, 5, 6, 8 (default 8)
• Metadata saving (PNG + JSON with both legacy and new fields)

Installation

Using ComfyUI-Manager (recommended):

• Search for “Mflux-ComfyUI” and install.

Manual:

1. cd /path/to/ComfyUI/custom_nodes
2. git clone https://github.com/joonsoome/Mflux-ComfyUI.git
3. Activate your ComfyUI venv and install mflux 0.10.x:
   • pip install --upgrade pip wheel setuptools
   • pip install 'mlx>=0.27.0' 'huggingface_hub>=0.24'
   • pip install 'mflux==0.10.0'
4. Restart ComfyUI

Notes:

• requirements.txt pins mflux==0.10.0; pyproject uses mflux>=0.10.0
• MLX < 0.27.0 will show a UI hint; upgrade is recommended for stability/perf

Nodes

Under MFlux/Air:

• QuickMfluxNode (txt2img/img2img/LoRA/ControlNet in one)
• Mflux Models Loader (pick saved models under models/Mflux)
• Mflux Models Downloader (fetch curated or third‑party models)
• Mflux Custom Models (compose and save custom quantized variants)

Under MFlux/Pro:

• Mflux Img2Img
• Mflux Loras Loader
• Mflux ControlNet Loader

Usage tips

• LoRA + quantize < 8 is not supported → set quantize to 8 when using LoRAs
• Width/Height should be multiples of 8
• dev model respects guidance; schnell ignores guidance (safe to set)
• Seed −1 randomizes each run. Presets are provided for size/quality.

Paths

• Quantized models: ComfyUI/models/Mflux
• LoRAs: ComfyUI/models/loras (tip: create models/loras/Mflux to keep them tidy)
• Full models cache (HF): ~/Library/Caches/mflux (or system cache)

Workflows

See the workflows folder for examples (txt2img, img2img, LoRA stack, ControlNet canny, and integrations). If nodes are red, use ComfyUI-Manager “One‑click Install Missing Nodes.”

Acknowledgements

• mflux by @filipstrand and contributors: https://github.com/filipstrand/mflux
• Some code structure inspired by @CharafChnioune’s MFLUX-WEBUI (Apache‑2.0) with license notes in the referenced areas

License

MIT