File	Size	Component
`dit.npz`	16.1 GB	DiT (8B, 36 layers) — pre-transposed for MLX
`vae.npz`	100 MB	FLUX.2 VAE decoder — pre-transposed for MLX
`bn_stats.npz`	tiny	Batch norm running stats for latent denormalization
`config.json`	tiny	DiT architecture config

Conv2d weights are pre-transposed from PyTorch NCHW to MLX NHWC format. No conversion needed at runtime.

Usage

python

from ernie_image import ErnieImagePipeline, TextEncoder

te = TextEncoder.from_pretrained()
pipe = ErnieImagePipeline.from_pretrained("treadon/ERNIE-Image-Turbo-MLX")

emb = te.encode("A vibrant manga comic about a cat and a dragon")
img = pipe.generate(text_embeddings=emb)
img.save("output.png")

Benchmarks (1024x1024, 8 steps, M4 Pro 64GB)

MLX vs PyTorch/MPS

Pipeline	Total	Per Step
PyTorch/MPS (diffusers)	137.0s	17.1s/step
MLX (this repo)	134.2s	16.0s/step

Breakdown

Component	Time
Text encode (PyTorch)	0.1s
Denoise (MLX)	128s
VAE decode (MLX)	6s
Total	~134s

Code

github.com/treadon/mlx-ernie-image

Follow @treadon on X for more ML experiments

Base Model

baidu/ERNIE-Image-Turbo — 8B DiT + Mistral-3 text encoder + FLUX.2 VAE. Apache 2.0.

More from me

For other projects and writeups, see riteshkhanna.com, follow @treadon on X, or treadon on Hugging Face.