arraypress/imagine-flux2-klein-4b

FLUX.2 [klein] 4B — Core AI (float16)

Pictures from words, on Apple silicon: Black Forest Labs' FLUX.2 [klein] 4B converted to a Core AI diffusion bundle, 1024 × 1024, 4 steps. It also edits an existing picture (image-to-image) with the two extra parts in this repository.

Credit and licence

• Model: FLUX.2 [klein] 4B, Black Forest Labs, Apache-2.0 — text encoder (Qwen3) and VAE included in that release. See LICENSE and NOTICE. Black Forest Labs encourages deployers to add filters for unlawful content.
• Conversion: Apple's official recipe, apple/coreai-models (BSD-3) at commit 52dbd93, unmodified: uv run coreai.diffusion.export flux2-klein-4b --platform macOS --single-function --resolution 1024 --compression none (float16). Not 4-bit — the recipe's default 4-bit compression was measured by others to fail parity, and the multi-function default needed over 55 GB of scratch space to compile.

This repository only converts and hosts the weights; all credit for the model belongs to Black Forest Labs.

Files (flat — the folder is the bundle)

	size	needed for
metadata.json, tokenizer/, vae_bn_mean.npy, vae_bn_var.npy	small	everything
TextEncoder.aimodel	5.8 GB	everything
Transformer.aimodel	7.2 GB	text-to-image
VAEDecoder.aimodel	95 MB	everything
Transformer_img2img_full.aimodel	7.2 GB	editing only
VAEEncoder.aimodel	66 MB	editing only

Apple's pipeline loads without the two editing parts and simply reports that it cannot edit, so a text-to-image install is 13 GB and editing adds 7.3 GB. An .aimodel is a build artefact; these are pinned by their published file hashes.

Shapes

shape	where	extra download	extra compiled cache
1:1, 1024 × 1024	top level	—	—
16:9, 1344 × 768	aspect-16x9/ — its own transformer and decoder, traced at 1344 × 768; shares ../TextEncoder.aimodel	7.3 GB	~7.3 GB
9:16, 768 × 1344	aspect-9x16/ — its own transformer and decoder, traced at 768 × 1344; shares ../TextEncoder.aimodel	7.3 GB	~7.3 GB
4:5, 896 × 1120	aspect-4x5/ — its own transformer and decoder, traced at 896 × 1120; shares ../TextEncoder.aimodel	7.3 GB	~7.3 GB

9:16 and 4:5 (added 2026-10-13) were exported with COREAI_IMAGE_HW=HEIGHTxWIDTH from the arraypress fork of apple/coreai-models at b667888 (tag 1.0.0-nonsquare.3), float16, no compression, transformer and decoder only. Measured on an M3 Max through imagine, warm: about 43–45 s for a one-picture command, the same seed giving a byte-identical file.

A Core AI transformer is traced for one shape. Measured before choosing this layout: a two-function transformer (1:1 + 16:9) shares one weight copy on disk (7.2 GB either way), but Core AI compiles every function at the asset's first load and keeps a full compiled copy of each (14 GB), so every user would pay for every shape. One folder per shape is fetched only by whoever asks for it. 16:9 needs the non-square support in arraypress/coreai-models (a fork of Apple's 1.0.0: image_width/image_height in metadata.json); 1:1 runs on Apple's pipeline as is. Editing is 1:1 only.

Measured (macOS 27, M3 Max, 36 GB, GPU, 1024 × 1024)

Measure	Result
per picture, model loaded	19.1 s
first picture in a fresh process (load + compile from the OS cache)	46.6 s
peak memory (all parts resident / loaded as needed)	8.6 GB / 4.6 GB
edit (image-to-image), one picture	84 s, 4.2 GB
same seed, run to run	byte-identical

Of five candidates measured on the same eight prompts, it was the only one to spell a short sign ("OPEN") correctly. Longer text comes out misspelled and the model invents lettering on things like cover art — the tools built on it append "no text, no letters, no typography" by default, which removed it, and add real text afterwards. It does not fit an iPhone's per-app memory.

The first load compiles the model for the Mac's GPU; macOS caches that per program under ~/Library/Caches/coreai-cache, about the model's own size.

Use

With the imagine command-line tool, which fetches this on first use:

imagine "a red bicycle against a blue door" -o bike.png
imagine --edit bike.png "make it night" -o night.png