gwenn-ha-dev/Siliconed v0.15

Siliconed 0.15

on GitHub

Compact by default

• The Compact is now preselected, and marked recommended, when you install a model on a Mac with
  more than 8 GB (the Light stays preselected on 8 GB). On an M1 Pro it renders as fast as the Standard or faster, takes half its disk space, and has the Neural Engine on by default. A model already
  installed keeps its version; the three stay one click apart.
• Full precision costs almost nothing. The Standard, the publisher's bf16 weights computed in
  fp32, takes the Compact's time within 4 % on Z-Image and Anima, and 6 to 10 % more on Krea 2 and
  Qwen-Image-2.1, whose weights do not fit in memory at full precision and are read again from disk
  at each step. Every figure: Numbers.

Larger images, lower peaks

• No area ceiling any more: this Mac's memory budget decides. The size picker offers 1536² and
  2048², greyed with the Mac they need where the budget refuses them. Anima Turbo renders 2048² on a
  16 GB Mac, peaking at 5.4 GB.
• Lower peaks, the same pixels: weights are now read from disk a few tensors at a time instead
  of a layer at a time — 0.8 to 1.4 GB less for Krea 2, 1.1 GB for Z-Image, 0.5 GB for
  Qwen-Image-2.1.
• Fixed: Krea 2's memory crept up during the denoising (objects freed only at the end of the
  render).

Check our numbers yourself

• siliconed check <model> compares every stage of a model with the reference implementation,
  on your Mac, against goldens you write with the scripts of tools/oracle/ and whose fingerprints
  we publish. --neural-engine judges the Neural Engine's arithmetic, --variant compact|light
  the Compact and the Light against the reference run on the third party's published weights. The
  command ships inside the app (Contents/Helpers): it checks the very engine the app renders
  with, without Xcode. How, and what it costs: VERIFY.md.
• silicontrol measure repeats our timings and peaks with the app alone: a series of renders
  between two witnesses, the app's sampled peak memory, swap-outs, disk reads and each image's
  checksum, as a TSV. silicontrol set faster-rendering on|off switches Faster Rendering like
  Settings does.

Crop and save

• Crop a rendered image or an edit reference: C to enter, corner and side handles, rule of thirds
  while dragging, Return to accept, Escape to cancel; never under 512 px a side. The crop is a new
  history entry; the original stays.
• Save as PNG, JPEG or HEIC; the last format is remembered. JPEG and HEIC carry no metadata.

Engine

• Large formats split their attention between the GPU and the CPU at the share this Mac's
  measure finds fastest, the CPU heads' softmax in one NEON pass.
• A bf16-activation path for the denoiser's products, chosen per Mac by its measure, judged
  against diffusers in bf16; off on the M1 Pro, where it does not win.
• This Mac is measured again before the first render after the update, about a minute.