Repository navigation
FrameMind Studio 0.1.4 - UNSIGNED development preview
Pre-releaseFrameMind Studio 0.1.4
24 October 2026 · Windows x64 · unsigned development preview · culling local-1.6
This release runs locally on Windows. GitHub hosts source code and downloads;
no Studio application is deployed to Netlify or Vercel. Optional editor cloud
features and initial model downloads are documented in the READMEs.
Download the installer and SHA-256 checksum from
GitHub release v0.1.4.
The unsigned preview follows the existing development distribution policy.
Windows may show SmartScreen; do not disable Windows protection. A signed client
release and installation on clean Windows without WebView2 remain unverified.
Current verification and limits
- The final suite passed: 206 tests across 27 files, and TypeScript passed.
- Windows NSIS and both frontend builds passed for this version.
- Actual 512/1024 LaMa inference passed in isolated Chromium with unchanged
source and outside-mask pixels. Higher resolution is slower and is not a
guarantee of better photographic reconstruction. - Native WebView2 verified context comparison, Cancel/Escape, Apply and exact
Undo/Redo before the final brush additions. The later native Quick check
verified a real clone source, a PNG working copy and an unchanged original.
Native 1024 inference and a complete final adaptive UI run were not verified. - The local installed update preserved all 341 existing IndexedDB/localStorage
files by SHA-256. This was the preceding local 0.1.3-metadata build containing
these features, not a clean installation of the versioned 0.1.4 package. - Native export evidence belongs to 0.1.2; it was not repeated for this release.
Earlier culling calibrations below are dated historical checks, not a fresh
local-1.6 accuracy benchmark. No general photographic quality claim is made.
Updating existing culling results
Run culling again after upgrading to engine local-1.6. Older automatic results
need new measurements; manual K/R/X decisions remain authoritative. Region
selection is manual and never implies automatic face or subject recognition.
Adaptive retouch crops and optional higher resolution — 24 October 2026
- Distant disconnected mask regions receive independent crops, preserving their
own input scale. Nearby regions with overlapping surroundings are merged, with
at most 16 processed regions. Connected objects are not split into tiles. - Balanced, Fine detail and Wider context change the surrounding scene supplied
to the model. Optional comparison prepares up to three distinct crops from the
same unchanged source. Only Apply changes the photograph, as one history step;
Cancel and Escape discard the previews. Identical crops appear once. - Standard input remains 512 × 512. Explicit Detailed/LaMa mode offers Higher
detail at up to 1024 × 1024; small crops stay at 512 × 512. This reuses the same
pinned model weights, changing only input dimension metadata in memory. The
downloaded model and cached bytes are unchanged. Automatic and Fast use 512. - Higher resolution uses more memory and time and is not a guarantee of better
photographic detail. Source dimensions, lossless PNG working copies, immutable
originals and the existing brush/lasso workflow are retained. - CPU runtime initialization completes before a GPU trial. A GPU timeout restarts
retouch in a fresh CPU worker, avoiding concurrent initialization of the shared
WASM module. Pending histogram results are ignored after leaving adjustments.
During the adaptive-retouch stage, 201 unit tests and TypeScript passed, including recovery from GPU and
initialization timeouts, concurrent waiting callers, and a separate compute limit.
The synthetic adaptive check is node scripts/retouch-adaptive-smoke.mjs --hd;
see contributor instructions. The earlier culling
verification below does not validate this retouch change.
Sample brushes and conservative culling — 24 October 2026
- Clone copies real source pixels selected with Alt + click. Heal transfers that
sample's fine texture while adapting its low-frequency color to the destination.
Both retain the existing mask strength, lossless working copy and shared history. - Inpainting can optionally match a consistent smooth-boundary color difference.
Strong edges and clipped colors are excluded; correction is capped and remains
restricted to the repair mask. The option can be disabled. - In full-photo comparison, Mark the main subject selects a 128/256/384-pixel
region. Its original-scale sharpness replaces background sharpness in the score.
Weak subject detail is reviewed; it is not automatically rejected against a sharp
background. Regions persist, can be cleared, and never override manual K/R/X. - A miniature signature alone no longer proves a removable duplicate. JPEG/PNG
candidates require identical decoded source pixels, compared in bounded tiles;
RAW candidates require identical original bytes. Differences remain for review. - Preview and selected-subject measurements compare noise-corrected edge energy
at two scales. Conflicting noisy detail adds a review reason, never a sole reject.
Culling stays local and deterministic. Older automatic results need reanalysis;
manual decisions remain in force.
These changes have not been evaluated as a general photographic accuracy benchmark.
Full original photographs
- Click a culling photograph to open the entire original image file, replacing
the fixed 384 × 384 px detail viewer. Display uses the imported bytes without
another image conversion or compression. - The initial view fits the complete frame without cropping. Use the wheel or
100%, 200% and 400% controls to zoom, then drag or use arrows to pan. Physical
pixel scaling accounts for Windows display density. - Select up to four frames with
+to compare full photographs. K/R/X apply to
the active frame; 0 fits the image, 1 selects 100%, and Esc closes the viewer. - RAW still uses its embedded JPEG, not developed sensor data. The viewer uses
the import source rather than later editor adjustments or retouching.
Detail-guided culling in local-1.5
The nine fixed original-resolution samples now have a budget of up to four
additional samples guided by preview detail. The extra locations are measured
at original resolution and do not replace fixed coverage. They can veto a blur
rejection; fixed-grid measurements still determine scores and series references.
If only the extra locations show strong detail, the frame requires review with
an explanation of the difference between regions. Sharp background detail does
not prove that the photographer's subject is in focus.
No photographer's selection is turned into universal taste rules. Manual choices
retain priority, and an unexported photograph is not treated as a bad-photo label.
The read-only culling-selection-audit.mjs compares source basenames with exports,
uses the normal RAW preview/import path and keeps private reports under ignored output.
Verification on 21 October 2026:
- 155 tests and TypeScript passed. The browser UI check covers a sharp patch
between fixed samples, pervasive softness, manual overrides, profile changes,
confirmed removal and cancellation. Culling requests and page errors were zero. - Web and desktop frontend builds and the Windows NSIS build passed. The same
UI check passed in the installed WebView2 application using a separate profile.
The local update preserved all 341 existing IndexedDB/localStorage files by
SHA-256, and the installed executable matched the verified installer payload.
Package metadata remains 0.1.3; this does not update the public 0.1.3 download. - Two RAW sets contained 146 and 81 originals with 39 matching exported frames.
Reject suggestions among those exported frames changed from three to zero.
Keep counts were unchanged; ambiguous detail moved to review. This is one
photographer's selection, not a measure of general photographic accuracy. - The earlier 114-JPEG set changed from 94 keep / 13 review / 7 reject to
94 / 19 / 1. All 341 original-file SHA-256 hashes across the three sets remained
unchanged. Unexported photographs were never labelled as bad ground truth.
Earlier culling changes in local-1.4
Engine local-1.4 corrects cases where small residual edge energy prevented a
soft original from receiving a reject suggestion. A verified series can also
identify weak original-resolution detail hidden by downscaling. That path
requires matching source dimensions, enough informative samples, low noise,
independently weak candidate detail and a sharp reference. A sampled sharp patch
protects the candidate from that original-resolution rule.
Manual choices still take priority. Culling remains local and deterministic,
without AI, models, API keys or photo uploads. Suggestions do not delete original
files; removing rejects from the working set still requires confirmation.
Verification of the earlier local-1.4 update
- 147 tests across 22 files, TypeScript, web and desktop frontend builds, and the
Windows NSIS installer build passed. - Browser and native WebView2 checks verified identical original bytes, complete
frames without clipping, physical pixel scaling, drag/wheel zoom, four-frame
comparison and manual decisions across reruns. Browser DPR was 1.5; native DPR
was 1.25. Culling external requests and page errors were zero. - A read-only run on the same 114 sports JPEGs changed suggestions from
94 keep / 17 review / 3 reject to 94 / 13 / 7. All original SHA-256 hashes
remained unchanged. This is one calibration set, not a general accuracy claim. - Native checks used a separate WebView2 profile. The local installed update
preserved all 341 existing IndexedDB/localStorage files and matched the exact
NSIS executable payload. This records that local update, not a public release.
Bounded automatic samples can still miss a small sharp subject. Culling does not
recognize eyes, expression, the decisive moment or focus on a particular subject.
Scores are technical indicators rather than probabilities of photo quality.