Siliconed 0.11

• Enlarge ×4 (the button under the image, its context menu, the Image menu): macOS's own
  super-resolution, the kind of network Pixelmator uses. It sees only the image — no prompt —, so it
  sharpens edges, hair and texture without adding anything: 1024 × 1024 becomes 4096 × 4096 in a second
  or two, beside the original, which stays. The model is Apple's: macOS downloads it once, after
  asking. Needs macOS 26; silicontrol enlarge i7 does the same.
• The curtain compares an image with its enlargement, and any two marked images of the same
  proportions; zooming goes toward the pointer, on the canvas and under the curtain, to slide
  the line over a detail.
• Paste an image with ⌘V (outside the prompt): copied in Procreate, Photoshop, Preview or the
  Finder, it goes where a drop on the Images box would — an image to edit, or a subject to name.
• LoRA in the open: the stack sits right under the model; « Add » is a menu of the model's
  LoRAs; each one shows its strength's slider, its figure (double-click: back to 0.8) and a remove
  button — no more popover to open and close.
• Explore stands beside Generate: the same gesture, many small images instead of one.
• The rack says more, in fewer places: each box shows what its stage took as soon as it ends;
  hovering a stage's name says what it does (text encoder, DiT, VAE…); the memory a render needs is
  said once, under the rack; what keeps Generate greyed (« Write a prompt »…) is said above it.
• The queue scrolls past a few renders and shows the running one at its head: its image k/N in a
  batch, the time left, a stop that lets the rest go on. The diagnostic leaves the toolbar (Settings
  and the Help menu keep it).

Fixed

• A crash when marking images to compare, most likely with a large image still being read.
• PNG metadata always writes CFG scale: 1: Civitai's « Remix » no longer starts from its own
  default scale, which burnt the image.
• The Dock icon no longer bounces when the queue empties.

Siliconed 0.10.2

• A Settings window (Siliconed → Settings…, ⌘,) replaces the switches of the app menu: Developer
  Mode, Experimental (Named Subjects, Negative LoRA Strengths), the comparison's curtain or side by
  side, this Mac's capacity map, the silicontrol command (installed or not, and where) and the models
  folder, shown in the Finder; each with a sentence saying what it does.
• Negative LoRA strengths (Settings → Experimental): the LoRA sliders go from −1.5
  to 1.5, in the panel and in the exploration. A negative strength pushes the image away from what
  the LoRA learned: it is meant for « slider » LoRAs (age, detail…), trained to be used from −x to
  +x; on an ordinary LoRA it mostly damages the image, hence the switch. Off, a negative value is
  refused, with what to turn on. silicontrol add --lora name:-0.5 follows the same switch.
• A solver axis in the exploration (Developer Mode): Euler, Euler a, DPM++ 2M and UniPC side by
  side, same seed, same steps; silicontrol grid --x solver=euler,dpmpp-2m does the same. The
  sheet's subtitle names the solver and the ComfyUI schedule when they are not the defaults.

Fixed

• LoRA strengths typed with a point in the exploration (0.4, 0.8) were unreadable on a Mac
  set to a language that writes decimals with a comma, and « Explore » stayed greyed out. Both
  forms now read.

Siliconed 0.10.1

• Anima images credit Anima on Civitai. Drop an Anima render on Civitai and it links the model
  page, as Z-Image and Qwen-Image-2.1 renders already did: the PNG names Anima Turbo v1.1
  (circlestone_labs' own page), whose file is byte for byte the one the app installs. Compact and
  Light credit the same page.

Siliconed 0.10

A third model: Anima Turbo v1.1

• Anima Turbo v1.1, by CircleStone Labs and
  Comfy Org, joins Z-Image Turbo and Qwen-Image-2.1 in the app, without Developer Mode. A
  2-billion-parameter DiT (Cosmos-Predict2) that reads the prompt with Qwen3-0.6B and an adapter, and
  decodes with the Qwen-Image VAE. It makes illustration, painting and anime; no photorealism, by its
  publisher's choice. It knows artists by name, written with @ in front (@name), as its model
  card says.
• Its recipe: 8 evaluations, CFG 1. Text-to-image, image-to-image, LoRAs (kohya and ComfyUI
  layouts), the rectangular formats, the four solvers in Developer Mode; no editing by instruction,
  and no Detail: on this model it makes mosaic blocks, or nothing visible.
• Times: on an M1 Pro with 16 GB, Standard, seed 42, 8 evaluations, 512² in 12.1 s, 1024² in
  43.7 s, 1024×1536 in 75.6 s.
• Standard, Compact or Light: 5.65 GB, 4.10 GB (the DiT in GGUF Q8_0) or 3.31 GB (the DiT in
  GGUF Q4_K_M), the DiTs as vanes430 publishes them. The text encoder stays the publisher's bf16 in
  all three: no 8-bit version of it is published.
• Checked against the reference, stage by stage: at 1024², the denoiser's output is 1.43·10⁻⁵
  from an fp64 computation of the same evaluation, the fp32 reference 1.47·10⁻⁵; the VAE closer
  to the fp64 than the reference; a LoRA against ComfyUI's own merge.
• License: CircleStone Labs Non-Commercial for the model, which derives from NVIDIA's
  Cosmos-Predict2 (NVIDIA Open Model License). The images it makes may be used commercially,
  as its publisher's model card says. The app shows the license before the first byte is downloaded.

Qwen-Image-2.1: the recipe as published

• Every step at full size. Until now a 1024² generation took its first two steps at half size,
  a shortcut of ours; it is gone: the image is the Turbo's 8 published steps and nothing else. A
  1024² takes 134 s in the app (142 s from the command line) instead of ~120 s on an M1 Pro with 16 GB;
  1024×1536 232 s instead of 190 s. Edits are unchanged: they never took a step at half size.
• ComfyUI's schedule, to compare (Developer Mode): a « ComfyUI schedule » switch walks the noise
  levels ComfyUI uses for this model — the base model's, not the Turbo's — same seed, same steps.
  silicontrol add --schedule comfyui does the same.

Fixed

• An artist written @name in an Anima prompt was refused by silicontrol with Named Subjects on
  (« No subject is called @name »): a model that reads no images now keeps its @ as text, as the
  rack already did.
• The text encoder's time in the rack could read « ≈ 0.00 s » after a prompt read back from the
  render cache; only a prompt actually encoded now teaches it.
• The time left counted the running render as nothing until its first full step: the pace learned
  for that model and size now stands in.
• The rack's seed kept showing an old fixed seed after a silicontrol render with a random one; it
  now says Random.

Siliconed 0.9

Report My Configuration now measures the chip

• The diagnostic no longer renders a model. In about 15 seconds, without any model, it maps what
  your Mac's chip computes: the matrix products of the models' blocks on the GPU, on the CPU's matrix
  unit and on both at once, fp16 and bf16 products, attention through the engine's three kernels,
  weight widening on the CPU and on the GPU — each result checked against the exact one, computed in
  double precision. The JSON report says what the chip computes, and whether it computes it right.
• Fixed: with the Compact install, the diagnostic always said « differs from the reference ». It
  compared the 8-bit weights' output with the full-precision reference: the weights differed, not the
  chip.

The engine follows the map

• Once you have run it, Siliconed uses what it measured, from the next launch: the GPU and the
  CPU's matrix unit together on the large products, and at which share; the attention's heads split
  between the GPU and the CPU; the in-house attention kernel — each only if it is faster on this Mac,
  beyond the measurement's own noise, and only if it computed right. Until you run it, nothing
  changes. The sheet says what this Mac uses, and whether it applies now or at the next launch.
• After a macOS or Siliconed update, the map is set aside until you run it again.
• With a map, the same seed may give slightly different bits than without; the same Mac with the same
  map still gives the same bits.

Siliconed 0.8.2

Fixed: Z-Image Turbo follows its recipe

• 8 steps are now 8 full-size evaluations, as the model card says. Until 0.8.1, Siliconed ran
  7 (it followed an older diffusers schedule whose last step moved nothing), and at 1024² it
  computed the first two at half size. Both are gone; checked again against diffusers, stage by
  stage. The images come out cleaner, with finer detail.
• Z-Image Turbo 1024² now takes 113–119 s (118–120 s in Light), 512² ~34 s. The times
  published until 0.8.1 (82–85 s at 1024²) were for the shorter schedule; Siliconed's figures in the
  README are remeasured; its comparison with Draw Things now sets 8 evaluations against 8, Draw
  Things' figures being those measured on 13/10/2026.
• A Z-Image image made with 0.8.1 or earlier no longer comes out again to the bit; the same version
  on the same Mac still gives the same bits. Qwen-Image-2.1 is unchanged.

Timings are measured on a MacBook Pro M1 Pro with 16 GB.

Siliconed 0.8.1

Fixed

• The rack's last rows take the click again. Just above the column's footer, the LoRA row,
  + LoRA and Import… were drawn but did not respond until the column was scrolled. The
  footer now sits under the rack instead of over it.

PNG metadata Civitai reads

• Upload an image to Civitai and its settings show up. Beside its own keys, each PNG now carries
  the text Civitai reads: the prompt with a <lora:name:weight> tag per LoRA, then the steps, the
  solver, the seed, the size and the model. Z-Image Turbo and Qwen-Image-2.1 renders credit their
  Civitai page with no manual entry.
• LoRAs are credited too, by the sha256 of the file as it was published, kept when you import
  it. A LoRA imported with an earlier version is named but not credited: import it again to credit it.

Siliconed 0.8

Check for Updates

• In the app menu, on demand only. Siliconed asks GitHub for its latest release and compares it
  with the version running; if a newer one exists, it opens its page. Nothing is checked in the
  background, nothing is installed for you.

Solvers (Developer Mode)

• Euler ancestral, DPM++ 2M and UniPC, besides Euler, on Z-Image Turbo and Qwen-Image-2.1: the
  same steps, the same σ, another walk between them. Each one is checked against diffusers'
  scheduler of the same name. Euler, the default, gives the same images as before, bit for bit.
• Another image, the same time. UniPC may recompose a Qwen-Image-2.1 picture; the ancestral
  solver may flatten a background. Edits keep image 1 under all four. Previews and sketches show
  each solver's own estimate. In the rack, the history caption, the PNG and silicontrol add|grid --solver.

The numbers

• Remeasured through the app, every render queued by silicontrol: Z-Image Turbo 1024² in 82 to
  85 s, Qwen-Image-2.1 in 118 s, an edit with one image in 212 s, with two in 285 s; 0 swap-out.
  The README sets them next to the Draw Things engine on the same Mac.

Timings and memory are measured on a MacBook Pro M1 Pro with 16 GB.

Siliconed 0.7.1

0.7 was never released: 0.7.1 carries it, with the memory computation made exact in the cases it
missed.

Memory: a fixed budget

• Siliconed never takes more than its budget: three quarters of the Mac's memory minus 2 GB
  (10 GB on a 16 GB Mac, 4 on an 8 GB one). It no longer reads how busy the Mac is: the same render
  on the same Mac always gets the same answer, whatever else is open. What the other apps hold is
  theirs.
• Known before the first step. Each render computes the memory it needs from its format, its
  prompt, its images and its LoRAs, takes the fastest plan that fits in the budget, and checks its
  own peak against that figure. Nothing switches plans halfway through any more. The figure is an
  upper bound: a long prompt on Qwen-Image-2.1, a batch of several images and the app's own memory
  are counted, and a render whose need cannot be computed is refused rather than started blind.
• Greyed out ahead. A format or a model beyond this Mac's budget is greyed out before you click,
  with the Mac it would need: “needs a Mac with 24 GB”.
• Batches of Qwen-Image-2.1 no longer keep two copies of the model at once.
• Same images, bit for bit. Measured on the M1 Pro: Z-Image Turbo 1024² in 81 s, Qwen-Image-2.1
  1024² in 116 s; the largest edit (1024×1536, three images) peaks at 3.2 GB.
• The diagnostic report (schema 4) gives the budget, the need of each stage and the peak measured
  against it.

Named Subjects (Experimental)

• Nothing kept. Subjects now live in memory only, like the session's images: nothing is written
  to disk, they go when the app quits. A subjects folder written by 0.6 is deleted at launch.
• Images or Subjects. The box above the prompt holds either the images of an edit (1 to 3) or
  your subjects (as many as you like, the prompt citing 3 at most) — one or the other, never both.
• The name never reaches the model. “@alice” becomes “the person in image 1”: a name read as
  content gave Alice in Wonderland.

Timings and memory are measured on a MacBook Pro M1 Pro with 16 GB.

Siliconed 0.6

One Images box, above the prompt

• The images come first. On Qwen-Image-2.1, the box that holds the images to edit is now called
  Images and sits above the prompt: you lay the picture down, then say what to make of it. Its wires
  run past the prompt to the encoders that read them.
• One budget, counted once. With Named Subjects on (Experimental), the subjects live in the same
  box, under the edit's images: to the model, a subject you cite is one more image. The box's title
  counts both against the model's 3 images, and turns red past them.

Memory

• An idle Mac no longer refuses Qwen-Image-2.1 at 1024². A plain generation was charged the
  memory of an edit with 3 images, and the safety margin ignored what macOS frees by compressing the
  other apps' memory. Now a render is charged the peak measured for its own number of images, and
  the margin counts that compression: on the Mac of the report, from 2.9 to 4.5 GB available for
  the same desktop. The no-swap rule is unchanged: everything is still measured at the worst case.

Named Subjects (Experimental)

• Compare with the subject. “@ana as a western comic” keeps the picture's framing, as an edit
  does: the curtain under the image now compares the result with its image 1, whether it came from
  the Images box or from a subject (its frame, as it was sent).

Under the hood

• Superseded kernels and settings removed (a scalar attention kernel, a half-precision mode): one
  version of each computation lives, and it is the one that is checked. Nothing changes in the
  images.

Timings and memory are measured on a MacBook Pro M1 Pro with 16 GB.