Repository navigation
FrameMind Studio 0.1.5 - UNSIGNED development preview
Pre-releaseFrameMind Studio 0.1.5
24 October 2026 · Windows x64 · unsigned development preview
This patch addresses retouch startup, repeated model preparation and brush latency.
It preserves the existing models, repair resolution, mask pixels, local workflow
and culling engine local-1.6. Studio itself is not deployed to a hosting service.
- Ready models are reused across editor openings and photos. Concurrent callers
share one preparation; a worker crash clears readiness before retry. - Reading cached model bytes is shown as preparation, rather than a new download.
Cached files and completed downloads are checked against the pinned byte size. - Pinned MI-GAN/LaMa weights are read in the window scope and transferred into
the model worker. This avoids the desktop worker's cache/network path; model
parsing and inference remain in the worker. Recovery rereads cached bytes
instead of retaining a second full copy of the large model in memory. - Desktop supplies the packaged WASM binary and a Blob URL of the packaged
standalone module. This enables WASM threads without fetching their module
through the desktop origin. A forced one-thread recovery remains available. - MI-GAN starts on threaded WASM without repeated benchmark repairs. LaMa
retains GPU support, requests the high-performance adapter and keeps CPU fallback.
An initialized WASM runtime is reused instead of constructing an extra large
CPU model before GPU initialization. - The photograph has a separate canvas surface. Brush feedback and processing
animation no longer rescale the full original on every frame. Cursor movement
avoids React state updates, and drawing is limited to one animation frame. - Soft-stroke blur runs on the stroke bounds instead of the complete mask, at
most once per frame; releasing the pointer always commits its final pixels. - Context variants are planned and deduplicated before inference. The mask is
read and labelled once, and progress shows the actual number of distinct crops.
Verification: TypeScript and 23 targeted runtime, region and pixel-preservation
tests passed. The bounded synthetic comparison is
node scripts/retouch-performance-smoke.mjs; it compares the 0.1.4 brush with
this version on a 24 MP image, checks identical final masks, model reuse and real
MI-GAN inference without external model requests. Native verification is separate
from browser evidence. These checks are not a general photographic benchmark.
LaMa computation, particularly 1024-pixel input on CPU, still needs more time than
MI-GAN. This patch removes redundant work; it does not claim instant completion
on every device. Initial model downloads still require internet.
Measured in isolated Chromium on this machine: dispatching a burst of 48 brush
movements on the same 6000 × 4000 synthetic image took 2613 ms before and 2.14 ms
after. Final mask SHA-256 matched exactly. Blur covered 134,184,960 pixels before
and 85,618 after; processing animation drew the full photo 30 times in 500 ms
before and zero after. This is a bounded event-burst measurement, not a promise
of the same speedup for normal interaction or other hardware.
Preparation from cached MI-GAN bytes took 2004 ms. Five subsequent preparation
calls together took 0.11 ms and shared the same model session. Actual first
MI-GAN inference took 2451 ms on WASM; external requests and new model downloads
were zero. Native installation and GPU behavior are verified separately.
The final native WebView2 check passed without request overrides: cached-model
preparation took 3680 ms, returning to retouch took 322 ms, and the first complete
repair/save took 7743 ms. A subsequent repair/save took 6439 ms. Inference itself
took approximately 6244 and 5054 ms. Native GPU MI-GAN was measured separately
and was slower, so the release keeps MI-GAN on threaded WASM. No universal
instant-inference claim is made.
Native cloning, PNG saving, Undo, actual MI-GAN inference and model reuse passed.
The complete imported JPEG SHA-256 remained unchanged. There were no external
requests or repeat model downloads. The check used a separate profile and kept
all 345 normal workspace files unchanged. Native export and native LaMa/1024
inference were not retested for this patch.
The local 0.1.5 update also passed: the installed executable matches the exact
native-tested SHA-256, and all 345 existing workspace files remained unchanged.
Download: GitHub v0.1.5.
The installer remains unsigned, following previous preview releases. Do not
disable Windows protection. Earlier features and limits: 0.1.4.