Krea 2 Turbo in ComfyUI on a 16 GB Mac. Tested on an M1 MacBook Air 16 GB and an M5 Pro 48 GB.

Two workflows:

• Krea2-Turbo_Mac_16GB.json: Q4_K_M GGUF (7.5 GB) + fp8 text encoder (5.0 GB). For 16-32 GB Macs. Needs ComfyUI-GGUF.

• Krea2-Turbo_Mac.json: bf16, core nodes only. For 48 GB Macs.

Same recipe in both: 8 steps, cfg 1, euler, simple. Model links and the Mac gotchas are in note panels inside each workflow.

M1 MacBook Air 16 GB, 8 steps, nothing else running:

Image                Time     Peak memory   Max swap
cafe 832x1216        22m52s   14.6 GB       3.9 GB
carpenter 1024x1024  25m42s   14.7 GB       4.0 GB
street 896x1152      23m37s   14.6 GB       3.6 GB
lake 1216x832        27m38s   14.5 GB       4.7 GB

The gallery shows each M5 Pro bf16 example next to its M1 Q4 re-render, same prompt and seed. Same scene every time; Q4 moves smaller details around. The cover image was rendered on the M1 too.

Seed hunting on 16 GB: queue a few seeds at 8 steps and walk away. Keep batch_size at 1 (batches break on MPS; separate queued jobs are fine). Re-render a keeper at more steps only if it needs more texture.

Install (16 GB):

1. Install ComfyUI-GGUF from ComfyUI-Manager.

2. Download krea2_turbo-Q4_K_M.gguf (vantagewithai/Krea-2-Turbo-GGUF) into models/diffusion_models, qwen3vl_4b_fp8_scaled.safetensors (Comfy-Org/Qwen3-VL) into models/text_encoders, qwen_image_vae.safetensors (Comfy-Org/Krea-2) into models/vae.

3. Drop in Krea2-Turbo_Mac_16GB.json. Close browsers and heavy apps before rendering.

Full recipe, every MPS trap and the bf16/fp8 notes: github.com/Bambushu/krea2-turbo-mac

Credits: Krea AI (Krea 2), vantagewithai (GGUF quants), city96 (ComfyUI-GGUF), Comfy-Org (repackaged weights).