Serika PhotoEdit 0.0.2

Added

• AI background removal and subject selection on this computer: BiRefNet (MIT) through ONNX Runtime, loaded only when present, with DirectML GPU acceleration when that runtime is installed. Preferences › AI picks the engine, shows its status and downloads the model with checksum verification.
• A low-resource Fast engine remains the default; Select › Subject (AI), Layer › Remove Background and Layer › Remove Background (AI) choose explicitly.
• Live MCP sessions: SerikaPhotoEdit --mcp --attach --mcp-root <folder> lets AI agents edit the documents open in the window, with every change on screen and in the undo history. Serika starts if it is not running, the title bar shows when an agent is connected, and Preferences › AI can turn agent access off.
• MCP protocol 2026-07-28 (stateless requests, server/discover, cache hints, version errors) alongside the 2025 handshake for existing clients.
• MCP tools for subject selection, background removal, layer masks, image placement, duplicating, deleting, grouping and restacking layers, single-layer previews, history, AI status and a documented action catalog.
• Photoshop-style Layers panel: an eye column (click, drag down to toggle many, Alt-click to solo), lock and footer icons, folder icons for groups, checkerboard thumbnails and a type glyph for text layers.
• Adjustments panel icons for the sixteen Photoshop adjustments, with the rest under More adjustments.
• Redrawn tool icons with a distinct glyph for every tool and Photoshop's corner triangle on toolbar slots that hold several tools, plus overlapping foreground/background colour chips.
• Optional GPU (OpenGL 3.3) adjustments and Gaussian blur with CPU fallback, a persistent preference and device diagnostics.
• LittleCMS ICC CMYK soft proofing, intents and black point compensation, paper simulation, gamut warnings, and 8/16-bit CMYK TIFF and ink separations. Layer editing remains RGB.
• Editable mixed typography, paragraph metrics, text on paths and shared render/outline layout.
• Native stdio MCP server with workspace access controls, isolated documents, previews, layers, native saves, atomic actions and undo/redo.
• Home screen featuring the PhotoEdit mascot, with matching Windows, Linux and macOS application icons.
• Multi-layer alignment, distribution, grouping, duplication, deletion and stack ordering, with stable selections and undo.
• Editable embedded Smart Object documents, linked placement/reload/relink, replacement and export, with retained transforms, masks and Smart Filters.
• Brush dynamics, deterministic stroke spacing, pressure/tilt controls, stroke previews and Serika brush preset import/export.
• Draggable RGB and per-channel Curves, histogram overlays, numeric point editing and curve presets.
• Pattern Fill layers and five gradient styles with multiple colour/opacity stops, plus an editable Gradient Map ramp.
• MCP layout tools for precise designs: text laid out in a box (wrapped or kept line by line, aligned by its ink and centred optically, optionally shrunk to fit), measure_text, align_layers, check_layout and full-resolution region previews, plus layout rules in the server instructions. inspect_document reports every layer's ink bounds, and text layers can be restyled.
• Super resolution like Photoshop's: Image › Super Resolution enlarges the whole document and Layer › Upscale Layer enlarges one layer about its centre, with AI (Real-ESRGAN General x4v3 or x4plus, BSD-3, run locally through ONNX Runtime on 128 px tiles, cancellable with progress) or Preserve Details (Lanczos-3 with edge sharpening, instant). Preferences › AI downloads and verifies either model.
• MCP upscale_layer and upscale_document, place_image resample (smooth, preserve-details, ai) and replace_layer_id to re-render a placed image in place (keeping its mask and stacking), upscale models in ai_status, and a warning when an image is enlarged with plain smoothing.
• serika_iconsheet and serika_psdcheck developer tools, scripts/psd_inspect.py and scripts/release.py.

Changed

• Preferences redesigned: a searchable sidebar with icons, settings grouped into cards with descriptions, on/off switches, live theme previews, model and runtime status badges, and Restore Defaults. Placeholder pages and settings that had no effect were removed.
• ONNX Runtime is shared by every AI feature, so subject selection and super resolution load it once.
• Select subject / Remove background / Add mask is a floating bar near the bottom of the canvas and is off by default (View › Contextual Task Bar).
• Essentials uses Photoshop's three panel stacks, with Properties as a tab beside Adjustments, so Layers keeps its room on short screens. Workspace layouts saved by earlier builds are reset once.
• Icons are drawn as vectors at the display's scale, so they stay sharp at 125 % and 150 %.
• The PSD import report lists only features imported as pixels, once per layer, instead of every preserved metadata block.
• Serika brand refresh: violet accents, plum and lavender surfaces, and bundled Outfit and Onest typefaces across all four themes, with visible keyboard focus and neutral image surrounds.
• Multi-layer Move and keyboard nudging preserve selections and unlinked masks, with gesture cancellation and atomic undo.
• The version is defined once in CMake and generated into the application.

Fixed

• PSD files from Photoshop opened upside down with empty groups, because layer records were read top-to-bottom; PSDs saved by Serika had the mirror-image problem in Photoshop. Files saved by earlier Serika builds still open correctly.
• Hidden layers could not be shown again: the dark theme drew nothing for an unchecked visibility box.
• Dragging layers in the Layers panel could reorder unpredictably or nest layers inside non-group layers.
• Remove Background used the whole document instead of the target layer, so a placed photo kept its background.
• The Adjustments panel's two-letter labels were ambiguous (three buttons read "Au").
• Editing documents with many layers was very slow: every change re-rendered each layer through full-canvas images. Plain, masked and text layers and full-strength pass-through groups now render directly, with identical pixels; a 63-layer cover dropped from about 5 seconds to 0.3 seconds per change.
• Adding to, subtracting from or intersecting a selection on a large document took about a second per operation; long MCP action lists froze the window until Windows closed it as not responding.
• MCP place_image ignored a single width or height instead of scaling proportionally.
• The History states preference only applied to documents that were already open.