Still on Rhossili's sand: 0.31 is the engine of 0.30 made faster, the Compact first, and held closer to its memory limit.
From 0.30, download 0.31 from the site. 0.30 does not tell you about it: on some Macs its « new version » panel closed the app at launch, so 0.30 no longer sees any new version. 0.31 fixes the panel, and from it on the app reads every published version again.
Faster
- The Compact computes int8 × int8 on the GPU. Core AI's own integer product now multiplies the Compact's published codes with the activation brought to int8 row by row, the recipe of its int8 reference; the Neural Engine's rows keep their W8A16, on a copy of the codes laid out for it. A step of the DiT at 1024² on the Mac mini M5 Pro is 10 to 40 % faster than in 0.30, for the five models. The Compact now buys time as well as memory, where 0.30 said it bought only memory (Numbers).
- The decoder keeps the GPU busy: its bands run two calls in flight instead of one at a time, while the CPU prepares the next. The VAE at 1024² on the Mac mini: Z-Image 1.25 → 0.89 s, Qwen-Image-2.1 1.67 → 1.26 s, Anima and Krea 2 0.77 → 0.48 s, MOSS-Image 1.32 → 0.92 s, the same pixels to the bit.
- A size never rendered decodes at full pace from the first image: its bands are sized from what the other sizes held on this Mac, where 0.30 fell back to the narrowest bands (4.5 to 8 s for the VAE at a new size).
- The Neural Engine is kept only where it helps: the engine also times each model on the GPU alone and keeps the faster. On the Mac mini, Anima and Krea 2 now render on the GPU alone. The first renders of a model at a new size each try one way before the fastest is kept.
- Times on the Mac mini M5 Pro, 1024², 8 steps, the model loaded. The Compact: Z-Image 23.8 → 18.3 s, Anima 10.2 → 9.1, Krea 2 46.9 → 30.5, MOSS-Image 24.7 → 19.7, Qwen-Image-2.1 32.2 → 24.0. The Standard: Z-Image 23.8 → 23.1 s, Anima 7.9 → 7.9, Krea 2 43.2 → 39.6, MOSS-Image 25.7 → 23.0, and Qwen-Image-2.1 29.7 → 32.4, slower. The 16 GB M1 Pro's are still 0.30's (Numbers).
The memory limit, held tighter
- Never past the limit at a height never read. In 0.30, a DiT's calls at a height the Mac had not measured yet were estimated by scaling another height by its rows, which could fall short (by up to 650 MB on the Mac mini); a decoder's narrower bands too. Both are now bounded from what the functions hold, never below a measure.
- The limit holds from one render to the next. Memory macOS still bills after it is freed is handed back before each stage is planned, so that the third render of a session plans like the first.
- The decoder is judged before the DiT: a decoding that cannot fit is refused before the minutes of the DiT, not after them. If a decoding is stopped anyway, the DiT's result is kept and decoded once more, without running the DiT again.
- A larger image is judged from the smaller ones. At a size larger than any this Mac rendered, what each stage will hold is predicted from the sizes it did, and a render that cannot fit is refused up front, with the number it is short by. At a size with nothing to go on, the render goes ahead under the limit's watch.
- A form that once stopped at the limit is tried again under a larger one: raising the memory limit in Settings lets it run again, where 0.30 never took it back.
- An update keeps what this Mac learned of a stage's memory, as long as the stage computes the same way.
Fixed
- Under the Neural Engine, the first step of a render was slightly off (2.8 % from the GPU alone): a buffer was read before the GPU had written it. Fixed.
- A MOSS-Image fine-tune whose file names its tensors under a prefix imported, then would not render; it renders now, as the other four models' did.
- A Qwen-Image-2.1 that renders only black or broken images after an update from 0.21 (issue 17): the files a model's versions share (tokenizers, VAE) are now checked once against their published bytes; a damaged one is removed and downloaded again on its own.
- A render that goes wrong says where: a non-finite result turns the DiT's box red, not the VAE's, and is never decoded; an image lost among several seeds is said, by its seed and its stage, the others kept, and the progress still reaches its end.
- Krea 2 reads LoRAs under diffusers' names again (issue 17), with the ones that adapt the timestep and the last layer: the 4-step distillation LoRA that 0.21 applied and 0.30 refused renders again.
- A long prompt pasted no longer spills below its box (issue 18), with scroll bars always shown.
- Reuse These Settings and Render Again bring back the image's own edit images, and none for an image made from noise, instead of keeping the Prompt Lab's.
- A series' grid follows the window as it shrinks as well as grows; Settings no longer leaves half its window empty under the Prompt Lab; the Prompt Lab's image tiles sit side by side; the edition's name stays readable over a pale sky; a grid keeps a LoRA's name after the rack changes model; the detached Chain panel keeps its minimum size.
The app
- The Neural Engine switch replaces Precision (Settings › General), on by default: it never chose a precision, only whether the engine may use the Neural Engine. 0.30's Exact becomes the switch off.
silicontrol set neural-engine on|off;siliconed check --neural-engine off|on|ROWS. - The seed: a Fixed box beside the field and the dice. Off, the field shows greyed the seed the last render drew, selectable to copy, and « — » before the first; checking Fixed keeps it. The Seeds bubble says which seeds it renders (3 seeds: 9402 to 9404).
- ⌘0 gives a series' cells back to the window after you sized them by hand.
- Every greyed button says why when the pointer rests on it.
- The disk image opens on a background of its own.
silicontrol: a--flaga command does not take is refused by name, wherever it stands, and a prompt that begins with--goes after a lone--;addandgridfill the rack as a click does, without the series it held before,gridwith the seed it drew;measurereports the Neural Engine switch and the memory limit (two more columns in its TSV);set precisionis gone, forset neural-engine.- Times from your Mac: an issue form takes what
silicontrol measureprints, the 8 GB Macs first (the form). - Render on another Mac: both Macs need 0.31.