Z-Image Turbo on M5 Max: existing run evidence, 20 October 2026 UTC

Hardware: 16-inch MacBook Pro, M5 Max 40-core GPU, 128 GB unified memory.
macOS 26.5.1 (25F80); ComfyUI v0.42.1, commit 63be2f9;
Python 3.13.16; PyTorch 2.14.1. Package versions: requirements-tested.txt.

A_bf16: full bf16 diffusion model and qwen_3_4b text encoder.
G_int8_fp8te: int8_convrot diffusion model and fp8_mixed text encoder.
Both: 1024x1024, 8 steps, res_multistep sampler, simple scheduler, cfg=1,
shift=3. Prompt and seed are recorded in each run JSON.
A cold/warm: 16.1 / 13.1 s. G cold/warm: 20.2 / 19.2 s.
A sampled peak process footprint: 27 GiB. G: 15 GiB.

run-*.json records submission-to-completion wall time, including polling
at one-second intervals. Cold means first image after loading the server;
warm follows with models loaded. No repeatability or hardware survey is
implied by these two runs per configuration.

mem-*.csv records epoch timestamps, process physical footprint from macOS
top, and system used memory from vm_stat. Column names say _gb but the
sampler converts to binary GiB. The process footprint is coarsely rounded
by top. Sampling sleeps one second after each measurement, so actual
spacing is longer. Use process footprint, not total system memory.
method/ contains the original measurement code with output paths made
relative. It is supplied for inspection; no models were rerun for this ZIP.
Server logs retain model loading and execution messages; ANSI color codes
and private installation paths were removed. Other run data is unchanged.

Public workflow/package download:
https://cdn.moda.app/voyager/blog/comfyui-on-mac/comfyui-mac-workflows-ceda63acef690dae.zip
Public one-prompt output comparison (not a quality-equivalence test):
https://cdn.moda.app/voyager/blog/comfyui-on-mac/fig-variants-131737586f19559a.webp
