A third of every render's peak memory was the MLX allocator's cache, and Phosphene never capped it. Now it does.
Peak memory down about 35%. A Q4 768x432 121-frame render went from 25.53 GB to 16.50 GB on a 64 GB machine - and got slightly faster doing it (71.45s -> 68.84s).
On a 16 GB Mac that same render sat at 101% of RAM. It now sits at 96%. That is the difference between swapping and running.
The frames are sha256-identical to the uncapped render. This is not a quality trade - it is memory that was never doing any work. The cap sizes itself from your machine (1/8 of RAM, floor 2 GiB, ceiling 8 GiB), and LTX_MLX_CACHE_GIB overrides it.
Also in 4.7.0
Motion Control. The Remix engine that turns your own footage into a new video was called "Control", and in eight months two separate people asked us to build the feature it already was. It now says what it does, and the words you would actually search - motion control, pose, depth, reference - are on the card without clicking anything.
H3 custom LoRAs now import from a file, including ComfyUI repacks. Contributed by @sahilkashyap64. The import reads the converter's own alpha metadata instead of assuming 1.0, so an adapter trained at alpha 8 / rank 128 no longer loads 16x hot. A 780 MB adapter costs +2.7 MB of RAM to import instead of +9.5 GB.
Fixes. Character training stopped recommending a preset that cannot carry a face. Image-to-video stopped substituting a reference image you never chose. And updates stop landing one click late: Pinokio reads update.js into memory and then runs it, so an update that fixed the updater always arrived a version behind. That script is now thin and delegates to a file read after the pull.
2,830 tests, 23 skipped.
Update from inside Pinokio.