mlboydaisuke/FLUX.2-klein-4B-CoreAI

FLUX.2 klein 4B — Core AI

Black Forest Labs' FLUX.2 [klein] 4B converted to Core AI for on-device image generation on Apple Silicon (macOS 27+), running on Apple's official diffusion runtime in apple/coreai-models.

FLUX.2 [klein] is step-distilled: 4 denoising steps at guidance 1.0 produce a full 1024×1024 image. It pairs a 4B flow-matching diffusion transformer (DiT) with an 8B Qwen3 text encoder.

Components

Component	Description
Transformer.aimodel	Flow-matching DiT (25 blocks), 1024×1024
TextEncoder.aimodel	Qwen3 text encoder (hidden states 9 / 18 / 27)
VAEDecoder.aimodel	Latent → 1024×1024 RGB image
VAEEncoder.aimodel	1024×1024 RGB image → latent (image-to-image)
tokenizer/, pipeline.json, vae_bn_*.npy	Sidecar assets (auto-loaded)

Weights are 4-bit quantized (int4, per-block, block size 32); compute precision float16. The full bundle is 4.0 GB — Transformer 2.0 GB · TextEncoder 1.8 GB · VAE 0.16 GB.

Usage

Sample app (easiest)

CoreAIImageGenMac — pick "FLUX.2 klein 4B", tap Download & Load, type a prompt, Generate.

Swift

import CoreAIDiffusionPipeline

let pipeline = try await Flux2Pipeline(from: modelURL)
let config = PipelineConfiguration(
    prompt: "a photo of a cat",
    stepCount: 4,
    guidanceScale: 1.0,
    schedulerType: .discreteFlow
)
let result = try await pipeline.generateImages(configuration: config) { _ in true }
let image = result.images.first!

Command line (zoo reference tool)

swift run -c release diffusion-runner \
  --model path/to/FLUX.2-klein-4B \
  --prompt "a photo of a cat" --steps 4 --guidance-scale 1.0

How it was converted

uv run coreai.diffusion.export flux2-klein-4b --platform macOS

Performance

M4 Max (128 GB): ~17 s for a 4-step 1024×1024 image (cold model load + 4 denoising steps + VAE decode). The distilled 4-step schedule means no negative prompt / CFG is needed (guidance 1.0).

License

Apache 2.0, inherited from the base model black-forest-labs/FLUX.2-klein-4B. The converted weights are redistributed under the same terms, with attribution to Black Forest Labs.