Krea2-MLX (Apple Silicon)

ComfyUI-Krea2-MLX

Run Krea-2-Turbo text-to-image locally on Apple Silicon inside ComfyUI, using Apple's MLX framework — fast and memory-efficient (quantized to fit a 32 GB Mac). Supports stackable LoRAs (official Krea / diffusers and ai-toolkit / PEFT formats).

The pure-MLX pipeline (krea2_engine/) is vendored directly in this repo — no external git dependency — and runs in-process, returning standard ComfyUI IMAGE tensors. It's a self-contained node group: load → (LoRAs) → generate → IMAGE, then use any ComfyUI node downstream (save, upscale, etc.). It does not expose a torch MODEL, so it does not plug into ComfyUI's native KSampler/ControlNet.

macOS on Apple Silicon (M-series) only. The MLX dependencies do not run on Windows/Linux. ~24 GB unified memory recommended for 1024².

Install

Via ComfyUI-Manager (recommended): search for Krea2-MLX and install; Manager runs requirements.txt, which installs the vendored pipeline's MLX deps into ComfyUI's Python.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/cthomasdesign/ComfyUI-Krea2-MLX
<ComfyUI-python> -m pip install -r ComfyUI-Krea2-MLX/requirements.txt

Restart ComfyUI.

Models

Put a transformer build in ComfyUI/models/krea2/ — the Krea2 Model Loader lists what's there and infers precision from the filename (mixed_4_8 / 8bit / bf16):

| Build | Size | Repo → file | |---|---|---| | mixed-4/8 | 9.8 GB | avlp12/Krea-2-Turbo-Alis-MLX-mixed-4-8 → transformer_mixed_4_8.safetensors | | 8-bit | 14 GB | avlp12/Krea-2-Turbo-Alis-MLX-8bit → transformer_8bit.safetensors |

On first generation the VAE + Qwen3-VL-4B text encoder (~9 GB) download automatically from krea/Krea-2-Turbo (cached in ~/.cache/krea2_alis_mlx). Put LoRAs in ComfyUI/models/loras/.

Nodes (category Krea2 MLX)

| Node | Purpose | |---|---| | Krea2 Model Loader (MLX) | Pick a build from models/krea2 → KREA2_PIPE (cached; stays resident). | | Krea2 LoRA (MLX) | Pick a LoRA from models/loras + strength → KREA2_LORASTACK. Chain several to stack. | | Krea2 Generate (MLX) | KREA2_PIPE + prompt/size/steps/seed (+ optional LoRA stack) → IMAGE. Progress bar + Cancel supported; NSFW safety_filter on by default. | | Krea2 Unload (MLX) | Frees the cached model and clears the Metal cache to reclaim memory. |

Minimal workflow: Krea2 Model Loader → Krea2 Generate → Preview Image. With LoRAs: Krea2 LoRA → (Krea2 LoRA → …) → Krea2 Generate. See example_workflows/.

Notes

• One 12.9B model (~18–20 GB with encoder/VAE) stays resident. On 32 GB, avoid loading large torch models simultaneously; use Krea2 Unload to free memory.
• The safety_filter runs the pure-MLX Falconsai/nsfw_image_detection classifier; the Krea 2 Community License (§4.2) requires reasonable content-filtering in deployments — keep it on there.
• Where required by law, disclose that outputs are AI-generated (License §4.3).

Performance

v0.4.0 optimizes the vendored engine — measured on an M1 Max 32 GB (mixed-4/8 build, 8 steps):

| | v0.3 | v0.4 | faster | |---|---|---|---| | 512², short prompt | ~103s | ~67s | ~35% | | 512², long prompt | ~104s | ~72s | ~31% | | 1024² | ~310s | ~263s | ~15% |

Repeat generations with the same prompt (e.g. seed sweeps) additionally skip the text encoder entirely via a per-prompt embedding cache.

What changed: native grouped-query attention; step-invariant conditioning hoisted out of the denoising loop; padded text tokens trimmed before the DiT (~30% of the sequence at 512²); dynamic-length text encoding; prompt-embedding cache. All changes are validated mathematically faithful to the reference computation (see tools/parity.py).

Seed reproducibility: the sequence-length changes alter bf16 kernel rounding order, so a fixed seed renders an equal-quality but not pixel-identical image vs v0.3 — the same class of change as an MLX or hardware upgrade. Within v0.4, generation is fully deterministic (same seed → same image). To re-render pre-0.4 seeds exactly, set KREA2_EXACT_LEGACY=1 (slower; restores padded encoding/attention and bypasses the cache).

License & attribution

Independent, unofficial — not an official Krea product or endorsed by Krea. The model is a Derivative of krea/Krea-2-Turbo under the Krea 2 Community License (see LICENSE); the MLX port is by avlp12 (Alis MLX), with LoRA support added in krea2-mlx-studio. The engine (krea2_engine/) is vendored here from that lineage — see NOTICE for the full attribution chain and what changed at each step. Model weights are downloaded by the user, not redistributed here. Node/engine code here: MIT (per upstream). NSFW classifier: Falconsai/nsfw_image_detection (Apache-2.0). By using the weights you agree to the Krea 2 Community License (commercial use requires < $1M annual revenue, or an enterprise license).