About
Run the latest state-of-the-art generative image models locally on your Mac in native MLX!
Table of contents
- 💡 Philosophy
- 💿 Installation
- 🎨 Models
- ✨ Features
- 🦄 Contributors
- 🌱 Related projects
- 🙏 Acknowledgements
- ⚖️ License
💡 Philosophy
MFLUX is a line-by-line MLX port of several state-of-the-art generative image models from the Huggingface Diffusers and Huggingface Transformers libraries. All models are implemented from scratch in MLX, using only tokenizers from the Huggingface Transformers library. MFLUX is purposefully kept minimal and explicit, @karpathy style.
💿 Installation
If you haven't already, install uv, then run:
uv tool install --upgrade mflux
After installation, the following command shows all available MFLUX CLI commands:
uv tool list
To generate your first image using, for example, the z-image-turbo model, run
mflux-generate-z-image-turbo \
--prompt "A puffin standing on a cliff" \
--width 1280 \
--height 500 \
--seed 42 \
--steps 9 \
-q 8
The first time you run this, the model will automatically download which can take some time. See the model section for the different options and features, and the common README for shared CLI patterns and examples.
Python API
Create a standalone generate.py script with inline uv dependencies:
#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.10"
# dependencies = [
# "mflux",
# ]
# ///
from mflux.models.z_image import ZImageTurbo
model = ZImageTurbo(quantize=8)
image = model.generate_image(
prompt="A puffin standing on a cliff",
seed=42,
num_inference_steps=9,
width=1280,
height=500,
)
image.save("puffin.png")
Run it with:
uv run generate.py
For more Python API inspiration, look at the CLI entry points for the respective models.
⚠️ Troubleshooting: hf_transfer error
If you encounter a ValueError: Fast download using 'hf_transfer' is enabled (HF_HUB_ENABLE_HF_TRANSFER=1) but 'hf_transfer' package is not available, you can install MFLUX with the hf_transfer package included:
uv tool install --upgrade mflux --with hf_transfer
This will enable faster model downloads from Hugging Face.
DGX / NVIDIA (uv tool install)
uv tool install --python 3.13 mflux
🎨 Models
MFLUX supports the following model families. They have different strengths and weaknesses; see each model’s README for full usage details.
| Model | Release date | Size | Type | Training | Description |
|---|---|---|---|---|---|
| Z-Image | Nov 2025 | 6B | Distilled & Base | Yes | Fast, small, very good quality and realism. |
| Krea 2 | Jun 2026 | 12B | Turbo (distilled) | No | Very good quality with a wide range of styles; good for creative exploration. |
| FLUX.2 | Jan 2026 | 4B & 9B | Distilled & Base | Yes | Fastest + smallest with very good quality and edit capabilities. |
| Ideogram 4 | Jun 2026 | 9B | Base | No | JSON-caption-native, typography-focused text-to-image generation. |
| ERNIE-Image | Apr 2026 | 8B | Distilled & Base | No | Single-stream DiT from Baidu. Vivid, high-contrast output. |
| Lens | May 2026 | 3.8B (+20B TE) | Turbo (distilled) | No | Dual-stream MMDiT from Microsoft with a GPT-OSS text encoder. Strong prompt adherence in 4 steps. |
| Ming-Image | Sep 2026 | 6.15B (+16B MoE TE) | Base | No | Design-focused (posters, cards, UI) with strong typography; outputs RGBA. |
| Boogu Image | Jun 2026 | 10B | Turbo (distilled) | No | DMD-distilled 4-step model with a photographic look and bilingual (EN/ZH) text rendering. |
| FIBO | Oct 2025+ | 8B | Distilled & Base | No | Very good JSON-based prompt understanding. Has edit capabilities. |
| SeedVR2 | Jun 2025 | 3B & 7B | — | No | Best upscaling model. |
| Qwen Image | Aug 2025+ | 20B | Base | No | Large model (slower); strong prompt understanding and world knowledge. Has edit capabilities |
| Qwen Image 2.1 | Sep 2026 | 7.1B (+8B TE) | Base | No | Single-stream block-causal DiT with a Qwen3-VL text encoder; 40-step guidance-free sampling. Multi-reference editing, LoRA, and RGBA. |
| Depth Pro | Oct 2024 | — | — | No | Very fast and accurate depth estimation model from Apple. |
| FLUX.1 | Aug 2024 | 12B | Distilled & Base | No (legacy) | Legacy option with decent quality. Has edit capabilities with 'Kontext' model and upscaling support via ControlNet |
✨ Features
General
- Quantization and local model loading
- LoRA support (multi-LoRA, scales, library lookup), including LyCORIS LoKr on FLUX.1 and FLUX.2
- Metadata export + reuse, plus prompt file support
Model-specific highlights
- Text-to-image and image-to-image generation.
- LoRA finetuning
- In-context editing, multi-image editing, and virtual try-on
- ControlNet (Canny), depth conditioning, fill/inpainting, and Redux
- Upscaling (SeedVR2 and Flux ControlNet)
- Depth map extraction and FIBO prompt tooling (VLM inspire/refine)
See the common README for detailed usage and examples, and use the model section above to browse specific models and capabilities.
🦄 Contributors
MFlux was originally created by Filip Strand in August 2024 and moved to this organisation in August 2026. It is maintained by:
 Filip Strand created mflux |
 Anthony Wu toolchain, CI, releases, mflux.web |
 plz12345 devops, Krea 2, Boogu |
 Mariano Abad ControlNets, training, Lens |
 Ian Scrivener community, CUDA, model builds |
Where the models and the main features came from, read from the merge history (gh pr view <n> --json author,mergedAt):
| Model or feature | Contributor | PR |
|---|---|---|
| FLUX.1 | @filipstrand | initial release, 2024-08 |
| Depth Pro | @filipstrand | #159 |
| Qwen Image | @filipstrand | #269 |
| FIBO | @filipstrand | #279 |
| Z-Image | @filipstrand | #284 |
| SeedVR2 | @filipstrand | #297 |
| FLUX.2 Klein | @filipstrand | #323 |
| FLUX.2 KV cache (klein-9b-kv) | @michaeltrefry | #426 |
| ERNIE-Image | @azrahello | #417 |
| Ideogram 4 | @omercelik | #433 |
| Krea 2 | @plz12345 | #453 |
| LyCORIS LoKr adapters | @JanGrohn | #422 |
| Fused-qkv LoRA loading | @deadmansahil | #459 |
| Boogu-Image | @plz12345 | #446 |
| PiD pixel-diffusion decoder | @azrahello | #490 |
| Z-Image Union ControlNet | @fxd0h | #482 |
| Krea 2 Raw, LoRA training, diffusers loading | @fxd0h | #462 |
| Lens (Turbo) | @fxd0h | #510 |
| mflux-capabilities, the machine-readable option contract | @fxd0h | #499 |
| Gradient checkpointing for training | @qruz-hq | #711 |
| Denoised prediction for in-loop callbacks | @IonDen | #729 |
| Qwen-Image-2.1 | @ivanfioravanti | #736 |
| Qwen-Image-2.1 reference editing, RGBA, prefix cache | @dreampuf | #741, #777 |
| Qwen-Image-2.1 masks, auto-mask, strength, verify | @flyingtimes | #749, #764 |
| PEFT and ComfyUI LoRA formats (Qwen 2.1, Z-Image) | @phplego | #756, #768, #772 |
| Text-prefix KV cache, Metal kernel, step cache | @murphymatt | #778, #779 |
| Ming-Image-0.1-Design | @joeynyc | #765 |
| load and generate entry points for UIs | @IonDen | #780, #785 |
| mflux.web UI packages | @anthonywu | #776 |
| CI gates, justfile, ty, the pypi environment | @anthonywu | #576, #590, #646 |
| Release process and release notes | @fxd0h | #685 |
| test_tiny fixtures for every model | @ianscrivener, @anthonywu | #611, #620, #599 |
Everyone else who fixed, tested and reviewed is in the contributor graph.
🌱 Related projects
Build a UI under mflux.web
We welcome independently distributed mflux.web.* implementations using Gradio,
FastAPI, FastHTML, or any other UI framework. Each project can choose its own
framework, dependencies, release schedule, and launch command.
mflux owns the parent initializer and inference implementation. Its package
path extends across installed distributions, including separate editable
checkouts. mflux.web is an implicit namespace package: Python combines the
mflux/web/ directories contributed by independently installed UI packages.
No intermediate distribution is required; neither core nor UI packages need to
depend on mflux-web. The mflux-web
distribution reserves the generic mflux-web name on PyPI; it does not provide
the shared namespace and does not need to be installed. Its optional demo is
just another child module, mflux.web.demo.
Your UI distribution owns a unique child, for example
src/mflux/web/example_ui/__init__.py. With uv_build, configure
module-name = "mflux.web.example_ui". Declare mflux and your chosen framework
as dependencies with versions your UI supports. The minimum mflux version for
split-directory installs must include this parent path extension.
Do not ship mflux/__init__.py, which belongs to core, or
mflux/web/__init__.py, which must remain absent for the implicit namespace.
Choose a unique child name to avoid collisions. PyPI distribution names can use
hyphens; Python child names must be valid identifiers, such as example_ui.
Reserving a PyPI distribution name does not reserve a Python namespace.
After installation in the same environment, consumers can use
import mflux.web.example_ui. Installing a UI does not automatically launch it,
discover applications, or mount routes. UI framework dependencies belong to the
individual UI distributions; this namespace support adds none to mflux.
Community applications
- MindCraft Studio — macOS app built on mflux by @shaoju
- mflux-paint — native macOS inpaint/edit app (pywebview), 16 models across edit/inpaint/text-to-image, mask painting, multi-seed batch, by @Amo643
- Mflux-ComfyUI by @raysers
- MFLUX-WEBUI by @CharafChnioune
- mflux-fasthtml by @anthonywu
- mflux-streamlit by @elitexp
- mlx-taef — TAESD/TAEF tiny-autoencoder live previews and low-memory FLUX decode for mflux, by @IonDen
- mlx-teacache — TeaCache step-skipping to speed up FLUX generation in mflux, by @IonDen
- MLXBits Image Studio - A native macOS Swift app for FLUX, Krea 2, Z-Image and more!
🙏 Acknowledgements
MFLUX would not be possible without the great work of:
- The MLX Team for MLX and MLX examples
- Black Forest Labs for the FLUX project
- Bria for the FIBO project
- Tongyi Lab for the Z-Image project
- Baidu for the ERNIE-Image project
- Ideogram for the Ideogram 4 project
- Krea.ai for the Krea 2 project
- Qwen Team for the Qwen Image project
- Microsoft for the Lens (Turbo) model, and Comfy-Org for the weights repackage
- inclusionAI for the Ming-Image project
- The Boogu team for the Boogu Image project
- ByteDance, @numz and @adrientoupet for the SeedVR2 project
- Hugging Face for the Diffusers library implementations
- Depth Pro authors for the Depth Pro model
- The MLX community and all contributors and testers
⚖️ License
This project is licensed under the MIT License.
1 maintainer
Metadata
1 maintainer
Release files for mflux 0.21.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| mflux-0.21.0.tar.gz | 1.1 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| mflux-0.21.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 2.5 MB
Release files / mflux-0.21.0.tar.gz
| Download URL | mflux-0.21.0.tar.gz |
|---|---|
| Size | 1.1 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
508ac1f0491a204968391eae0d507d77bd54941a1eea37d080ab3e33d53794f6
|
|
BLAKE2b-256 checksum How to use checksums |
04a9e9f741cc47906fd3a87b291949314c3bde211389d67eff50e878914500d7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 3, 2026.
Transparency logRelease files / mflux-0.21.0-py3-none-any.whl
| Download URL | mflux-0.21.0-py3-none-any.whl |
|---|---|
| Size | 1.5 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
b2e38fd3cd4e50d497cc226e01555b8fc7856181c9421bde2e043bc53ac12d2e
|
|
BLAKE2b-256 checksum How to use checksums |
d4b73e0e6577e90cc8bd13aecc50524e2cbeabaf548896de960fede2e9e82b5a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 3, 2026.
Transparency log
