You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This is a Core AI conversion of FLUX.3 [klein] by Black Forest Labs, under the FLUX Kommunity License (LICENSE.md): it may be used and distributed for non-commercial purposes; only a Qualifying User may use its output commercially. You must use content filters or review output before distributing it, and disclose AI generation where the law requires. By agreeing you accept that licence.

FLUX.3 [klein] — Core AI

Black Forest Labs' FLUX.3 [klein] converted to Core AI diffusion bundles for Apple silicon, 4 steps — used by the imagine tool. Any shape and size up to native 4K, and editing with up to ten pictures.

Two bundles, the same model:

	weights	download	use it for
`full/` (`imagine --flux3`)	bfloat16	17 GB	text — signs and handwriting spelled right
top level (`imagine --flux3-small`)	4-bit	5.4 GB	drafts — prints signs, garbles handwriting

Files

Each size — full/ (bfloat16, imagine --flux3) and the top level (4-bit, imagine --flux3-small) — is one set of files that draws and edits every shape:

file	what	bfloat16	4-bit
`Transformer_4k.aimodel`	text to image; image and text lengths variable	9.1 GB	2.9 GB
`Transformer_img2img_full.aimodel`	editing; image, text and reference lengths variable	10 GB	3.2 GB
`VAEDecoder_4k.aimodel`, `VAEEncoder_4k.aimodel`	any height and width up to 4096	0.16 GB	0.16 GB
`TextEncoder.aimodel`, `tokenizer`, `vae_bn_*.npy`	Qwen3-VL text encoder and sidecars	8.3 GB	2.5 GB

Any width and height in multiples of 16, sides up to 4096; imagine allows up to 3840 × 2160 (2880 × 2880 square), the largest measured — about 18 minutes at bfloat16 and 16 at 4 bits on a 36 GB Mac, the picture decoded in overlapping 1024-pixel tiles (a whole 4K decode alone ran the Mac out of memory). An edit takes up to ten pictures, each read whole as diffusers' FLUX.3 pipeline reads it — the edited picture at the output's size, every other one fitted under 1 megapixel at its own shape — each at its own position ids. Core AI compiles each variable function once, for every size.

Measured on an M3 Max (36 GB) through imagine: the variable transformer agrees with diffusers' own forward to cosine 0.999912 on saved reference inputs, exactly as the fixed 1:1 export did; a 2048 × 2048 poster with its title spelled right in 420 s; a 16:9 picture with its sign right in 116 s; a person dressed from a cardigan's and a beret's pictures (three in all) kept her face, 242 s; at 4 bits a 3:4 picture in 43 s.

The older files — 0.20's *_any.aimodel (up to 2048) and 0.17–0.19's Transformer.aimodel, VAEDecoder.aimodel and shapes/ — are kept for the imagine versions that pin earlier revisions; nothing current reads them.

Editing

Instructions are passed as written — the bundles do not include BFL's prompt upsampler — and imagine guide edit shows the formats (a full instruction, or a paragraph with a JSON table of boxes on a 0–1000 grid). Measured against diffusers' own pipeline on the same pictures: a mug recoloured, a notebook removed with a written instruction (a short "remove the notebook" left it, in diffusers too), a mug moved by a box record, a person dressed from garments' pictures.

Licence — read this first

FLUX Kommunity License v1.0 (LICENSE.md). The model and this conversion may be used and distributed for non-commercial purposes; a Qualifying User (under US$5M gross annual revenue, with affiliates) may also use the pictures commercially. Rights come directly from Black Forest Labs to you, not from us.
§2(e): you must either use content filters to prevent unlawful, unsafe or infringing output, or review output before distributing it — and disclose AI generation where the law requires. imagine writes the IPTC digital-source-type field (trainedAlgorithmicMedia) into every file.
BFL's Attribution Notice and our modification statement are in NOTICE. This is not an official Black Forest Labs product.
The text encoder is the language model of Qwen3-VL as shipped in BFL's release (Apache-2.0, LICENSE-QWEN3-VL).

Conversion

Apple's recipe, apple/coreai-models (BSD-3), extended for FLUX.3 in arraypress/coreai-models at tag 1.0.0-nonsquare.8; source revision 3373376.

bfloat16. float16 overflows (NaN, in diffusers too). Against diffusers in PyTorch the bfloat16 bundle's output agrees to cosine 0.99992.
4-bit: weights quantized per block of 32, symmetric, scales stored as float32 (Core AI's 4-bit GPU kernel rejects bfloat16 scales); norms excluded. The VAE is never quantized.
Recipe, as in diffusers' FLUX.3 pipeline: the text window is the prompt rounded up to a multiple of 80 tokens (up to 1040); padding sits at position 1 inside Qwen3-VL; image positions are scaled by 0.5; sigmas linspace(1, 1/n, n) shifted by 6.93, each step aimed at the next sigma.
4 steps where diffusers runs 16: on text, faces, hands and crowds, 4 steps matched 8 and 16.

Measured (macOS 27, M3 Max, 36 GB, GPU, 1024 × 1024, 4 steps, through `imagine`)

Measure	`full/`	4-bit
chalk menu "Today: tomato soup", seeds 21–24, spelled right	4 / 4	0 / 4
printed signs, labels, covers: six prompts, one seed	6 / 6	5 / 6
one picture, whole command	77–98 s	33–38 s
each further picture in the same run	25–32 s	—
peak memory footprint / most resident	6.1–7.7 GB / 23–26 GB	13.3 GB / 6.3–8.6 GB
first run (compiles for this Mac's GPU, cached per program)	≈16 min	≈30 s

At 4 bits the model still prints a sign but writes handwriting as near-letters ("yerunaoo soup"), so the bfloat16 bundle is the one to use for text. Faces, hands and crowds look the same at both.

An 8-bit bundle was measured and dropped: 8 GB, but ≈190 s a picture (Core AI re-attempts a Neural Engine compile on every load). Core AI also prints six compiler warnings to standard output on each load of the 4-bit bundle; imagine silences them.

Use

imagine model accept flux3          # once: shows these terms, records acceptance
imagine --flux3 "a chalkboard menu that reads 'Today: tomato soup'" -o menu.png
imagine --flux3-small "a lighthouse on a cliff, poster art" -o draft.png
imagine --flux3 "a wide street sign that reads 'Harbour Road'" --aspect 16:9
imagine --flux3 "a Lisbon travel poster with the word 'LISBOA'" --size 2048x2048
imagine --flux3 --edit person.jpg --ref cardigan.jpg --ref beret.jpg "Create a photograph of the woman from the first image wearing the cardigan from the second image and the beret from the third …"

Downloading needs a Hugging Face account that has accepted this repository's terms and a token (hf auth login); imagine says so with exit 3 when either is missing.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for arraypress/imagine-flux3-klein

Base model

black-forest-labs/flux-3-image-klein

Finetuned

(14)

this model