[lalibertemarc]

Description

Adds a second window that shows the image being edited, for multi-monitor setups (requested on Discord). The viewer has its own zoom and pan, independent of the editor. When zoomed in, it renders at higher resolution, so it stays sharp up to 100%.

Type of Change

• Bug fix
• New feature
• Breaking change
• Performance improvement
• Code refactoring
• Documentation update
• UI/UX improvement
• Build/CI or Dependency update

Changes Made

Opening the viewer

Either of these opens it. Both are hidden on Android.

• "Open in second window" in the editor's right-click menu
• Shift+F, a new rebindable action in the View keybind section (the key was unused)

There is no toolbar button, per maintainer feedback.

Viewer window

• Native window decorations.
• It opens on a monitor other than the editor's when there is one.
• It remembers its monitor, size and fullscreen state in viewer_window_state.json.
• F11 toggles fullscreen; Esc exits it.
• Closing the main window also closes the viewer, so the app still quits.

Zoom and pan

• Zoom goes from fit up to 100% (1 screen pixel = 1 image pixel).
• The viewer follows the Mouse Controls settings exactly like the editor canvas:
  • Zoom Speed Multiplier scales wheel and pinch zoom.
  • Mouse mode: the wheel zooms around the cursor; Shift+wheel pans horizontally, Alt+wheel vertically.
  • Touchpad mode: two-finger scroll pans, pinch zooms (with the editor's touchpad speed boost).
  • Left or middle drag pans.
  • A single click toggles zoom: to 100% when "Zoom photo to 100%" is on, otherwise to 2× (or the last click-zoom level). Clicking again returns to fit.
• F11 toggles fullscreen; Esc exits it.
• Switching to another image keeps the viewer's zoom and position, like the editor canvas. It is re-clamped to the new image's size.
• Before/after: the toggle compare mode (B) is mirrored in the viewer. The split and side-by-side modes stay in the editor; the viewer keeps showing the edited image, and the editor's original-layer renders are never sent to it.

Rendering

• wgpu renderer (Windows/macOS default): the viewer gets its own wgpu surface. Each frame is copied on the GPU into a texture the viewer owns. Viewer renders never write the texture that the main window and the white-balance picker read.
• Live updates: editor frames update the viewer live while sliders are dragged. Interactive frames rendered for a zoomed-in editor region (ROI) are skipped; the viewer updates when the slider is released.
• Sharp zoom:
  • When the visible area needs more pixels than the editor rendered, a viewer-only render is queued on the preview worker. This happens about 250 ms after zooming stops.
  • Editor jobs always run first.
  • The viewer has its own preview and GPU image cache slots, so it doesn't evict the editor's caches.
  • A lower-resolution frame of the same edit doesn't replace a sharper viewer render.
  • At fit, frames within 80–200% of the needed resolution are used as is, so the viewer doesn't re-render after every edit.
• No wgpu renderer (Linux, or renderer turned off): the same frames and renders are sent as JPEG and placed with a CSS transform.

Shared-code changes

• Display creation moved out of get_or_init_gpu_context into build_wgpu_display. The main window and the viewer both use it.
• WindowEvent::Resized is now routed per window. Before, resizing any window reconfigured the main surface.
• Window-state file read/write (read_window_state / write_window_state) is shared between the main window and the viewer.
• AppSettings::wgpu_renderer_enabled() replaces the duplicated cfg logic.
• The preview-resolution formula moved from useImageProcessing to calculateTargetResolution (src/utils/previewResolution.ts). The main editor calls it with the same inputs as before.
• clear_preview_slot replaces duplicated cache resets in load_image and clear_image_caches.
• The editor's wheel-gesture rules (zoom vs. pan per input mode and modifiers), click-zoom target and click slop moved to src/utils/canvasControls.ts (interpretWheel, clickZoomTarget, movedBeyondClick). The editor behaves as before, and the viewer uses the same code, so both canvases respond to the mouse settings identically.

Other

• New capability viewer.json for the viewer window label.
• Strings added to all 16 locales.

Screenshots/Videos

Context menu (please disregard the Edit in Topaz, only present on my end):

Testing

• These changes were tested locally by a human and confirmed to work.
• I haven't added any automated tests to the code because the codebase currently lacks a test suite.

Tested with the wgpu renderer on Windows:

• opening and closing from all three entry points
• resizing and drag-resizing the viewer
• live slider updates
• zoom, pan, and sharpening when zoomed in

@enrperes also tested it on macOS (see the comment below), with no issues found. The Linux / renderer-off JPEG path has not been tested by a human.

Test Configuration:

• OS: Windows 10 Home (19045)
• Hardware: AMD Ryzen 5 3600X, NVIDIA GeForce RTX 2070

Checklist

• My code follows the project's code style
• I haven't added unnecessary AI-generated code comments
• My changes generate no new warnings or errors

Additional Notes

Costs and limits:

• GPU memory: at 100% on a 24MP image, the viewer uses roughly 300–500 MB of extra VRAM: its own input texture, its frame texture, and possibly a larger processor. It is freed when the viewer closes. Nothing extra is allocated while the viewer is closed.
• Render time: a full-resolution viewer render takes about 0.3–1.5 s and can't be interrupted. If a slider drag starts during one, its first frame waits for it.
• Geometry and effects edits: the effects/transform caches are still single-entry. These edits rebuild once per cache slot. Colour-only edits stay GPU-only in both slots.
• macOS: the viewer surface is created on the main thread, which NSView requires.

AI Disclaimer:

Please state the involvement of AI in this PR:

• This PR is created by an AI agent
• This PR is mostly AI-generated but edited/merged together by a human
• This PR was handwritten with AI assistance (spell check, logic suggestions, error resolving)
• This PR contains only blood, sweat, and coffee (AI-free)

🤖 Generated with Claude Code

[enrperes]

I tested this on macOS and it works great! I haven't encountered any issues.

[CyberTimon]

LGTM! Thanks so much for your work on this, much appreciated! 🥇