devin-lai/flux-3-image-klein-mlx

FLUX.3 Image Klein on your Mac

Local image generation and editing with MLX. Five model choices. Download only the one you need.

Q8 quality, 16 steps. These are existing benchmark outputs, resized for this overview. Full prompts, seeds, inputs and original PNGs.

An independent Apple Silicon port of Black Forest Labs' distilled FLUX.3 Image Klein, published by Devin Lai. Includes the inference runtime, ready-to-run weights, exact presets, benchmark data and original comparison images. The image transformer has 5.6B parameters; the text encoder is separate.

Start with Q8 quality. On the tested MacBook Pro M5 / 32 GiB, its 16-step preset measured 38.75 s across four workloads, 7.46 GiB maximum process footprint, and 4.16× the speed of the strict MLX FP32 reference. Its 11-step preset measured 29.69 s in a later round. These are workload aggregates, not promises for every image size or Mac. Measurement details.

Choose your download

This repository holds the runtime, documentation and gallery. Each linked variant repository contains one complete model configuration, including its text encoder, tokenizer, VAE and image-editing components. No other variant is needed.

Model	Download¹	Local time²	Peak memory²	Choose it for
Q8 quality	12.73 GiB	38.75 s	7.46 GiB	Recommended balance of fidelity and storage; Q8 transformer + BF16 text
Q8	9.58 GiB	36.65 s	7.43 GiB	Smaller download; Q8 transformer + Q8 text
Mixed 4/8-bit	8.23 GiB	34.91 s	6.07 GiB	Lower memory, with larger visual changes
Q4	7.40 GiB	33.75 s	5.53 GiB	Smallest download, with stronger fidelity tradeoffs
BF16	17.40 GiB	40.34 s	12.47 GiB	Unquantized weight control

¹ Model components and weight licenses; small repository metadata is additional. Download size is different from runtime memory. ² 16 steps on M5/32 GiB; geometric mean of four per-workload median launch times, three fresh-process repeats per case. Peak memory is the maximum process lifetime footprint. Tested canvases range from 640 × 640 to 1024 × 1024 and include one reference-image edit. A minimum supported Mac memory capacity has not been established.

Q4 changed object counts in the typography example. Inspect the image comparisons and model selection guide when choosing a compact model.

Generate your first image

Use an Apple Silicon Mac and Python 3.12, preferably in a fresh virtual environment. Install the small runtime once; the generation command downloads only Q8 quality on its first run.

python3.12 -m venv .flux-env
source .flux-env/bin/activate
python -m pip install "huggingface_hub>=1,<3"

hf download devin-lai/flux-3-image-klein-mlx \
  --include "flux_klein_mlx/*" --include "pyproject.toml" \
  --include "LICENSE-runtime.txt" --include "examples/*" \
  --local-dir flux-klein-runtime
python -m pip install "./flux-klein-runtime[hub]"

flux-klein --download q8-quality --model-dir ./flux-models/q8-quality \
  --preset q8-quality-16 --prompt-file ./flux-klein-runtime/examples/teapot.json \
  --width 768 --height 512 --seed 1937 --out ./flux-output/teapot.png

After downloading, omit --download to run locally. Use --preset q8-quality-11 for the faster tested preset, or --download q8 --model-dir ./flux-models/q8 --preset q8-16 for the smaller Q8 package. Use a separate directory for each variant.

For editing, add --image /path/to/reference.png and supply a full scene caption describing the desired result. This CLI sends captions as provided; it does not run a prompt upsampler or bundle content filtering. Review generated outputs before sharing and follow the upstream license. Installation, offline use and editing guide.

Inspect the evidence

• Original image stacks: matched prompts and seeds, reference inputs, labeled comparisons and lossless PNGs.
• M5 / 32 GiB benchmark: 120 fresh-process measurements, latency ranges, memory, PSNR, SSIM and LPIPS.
• Presets: nine finalized presets across five configurations, including the exact 11/12/13-step grids.
• Results and limitations: recommendations and observed failure cases. Image similarity measures reference agreement, not aesthetic superiority.
• Runtime validation: ten packaging checks reproduced the recorded PNG bytes on the tested machine. The Hub download and release-packaging additions have separate software checks.
• Technical format and provenance: packed safetensors, component sharing, source revisions and inference arithmetic.

The measured FP32 reference upcasts the original BF16 checkpoint values and uses strict FP32 arithmetic. Accelerated presets also change arithmetic, kernels and VAE execution. The speed ratios compare configured pipelines; they do not isolate quantization gains or compare against Diffusers, CUDA, Core ML or another Mac application.

Model card and license

This is an inference distribution of the distilled checkpoint: text-to-image generation and reference-image editing. It repacks and quantizes existing weights; no new pretraining or fine-tuning dataset is introduced. Upstream training provenance, biases and model limitations remain those of the original model card. The runtime can accept up to ten reference images, but the local benchmark covers single-reference edits only.

The packed fluxmac-packed-v1 format requires this runtime. It is not a drop-in Diffusers pipeline or a complete Qwen3-VL language model. Conversion, calibration and experiment source code are outside this release.

Weights: FLUX Kommunity License. Text encoder assets: Apache-2.0. Inference runtime: Apache-2.0. Publishing the weights does not relicense them as unrestricted open-source software. See the attribution and modification notice. This is an independent derivative, without Black Forest Labs endorsement.

Share results and help improve Mac support

If this is useful, like the model you use and follow devin-lai for updates. Share outputs with the variant, preset, prompt, seed, canvas size and Mac configuration so others can reproduce them. Report installation or generation issues in this repository's Community tab; include the command and package versions, and remove tokens or private paths from logs.