Your library on another disk
- Settings moves the models to an external disk or a network share, opens a library already
there, or brings it back to the Mac. Each file is copied, read back and compared (SHA-256) before
anything switches; the old place is emptied last. Stopped or failed, the library is unchanged, and
a new attempt keeps the files already verified. A FAT disk, a read-only folder or another library
is refused before the first byte. (#2) - Renders never read the other disk. As soon as a model is chosen, what it reads is copied to
the Mac for the session (~/Library/Caches/Siliconed; several models kept, the least recently
used evicted, 5 GB always left free), and removed when the app quits — after a crash, at the next
launch. The same seed gives the same image as from the Mac's own disk. - What belongs to the Mac stays on it: the licenses you accepted, this Mac's measure, the render
cache. Unplugged, the disk is flagged: nothing renders until it is back, and an empty library never
takes its place. The remote control and the command line follow a moved library.
A model imported before its family
- Dropping a fine-tune before its family is installed (a Civitai or Hugging Face checkpoint of
Z-Image, Qwen-Image-2.1, Anima or Krea 2) now opens that family's installation, the same panel as
in the Models window: the license in full, the space, the text encoder and the VAE — never the
family's own DiT, since the file brings its own. Once installed, the file is imported, and the
license accepted there covers it: no second sheet at its first render. Until 0.12 the app only
said to install the family first. (#5) - A download refused by Hugging Face (HTTP 401 or 403) no longer sends you to a terminal to sign
in: no repository the app installs from asks for an account. The message now points at what
refuses — a proxy, a VPN, a firewall.
Fixed
- A render could come out entirely non-finite (Z-Image at 1024² under a LoRA, at certain prompt
lengths): Apple's fast fp32 matrix product sometimes writes invalid values into some output
buffers. Each buffer and shape is now checked once, and recomputed in a fresh buffer when wrong;
the same bits everywhere else, no measurable cost. - Detail on Qwen-Image-2.1: More and Most no longer change the person. Calibrated on the 6-step
LoRA the family used before 0.10, they weighed on steps where the Turbo still decides the face;
they now touch the last evaluation alone (More 1.3, Most 2): the same person, finer texture.
Normal is unchanged, to the bit. - The license sheet of an imported model says the image model is the imported file, and links
the licenses of the installed version's encoder and VAE — a Compact's or a Light's come from third
parties. - Two conversions of the same model at once (the app and the command line, say) no longer empty
each other's file in progress: the second one is refused. - Krea 2 renders' metadata name the Civitai page that carries the same bf16 file, so Civitai
recognizes the model. - Two engine settings never chosen by any Mac's measure (half-precision attention, GPU widening) are
removed; the default paths are unchanged, to the bit.