Scenery Layer Registration

Scenery Layer Registration

Use this skill for scenery/backdrop work where every mask and runtime asset must share one native coordinate space. In this repo the blue-hour waterfront source space is 2560x1440; masks and stackable layers must be full-frame, never cropped, and aligned by coordinates alone.

Non-Negotiables

• Treat registration as binary: aligned or not aligned.
• Never fix registration by moving the subject until it looks acceptable.
• Never crop a layer/mask to its visible content. Runtime stacking depends on exact same-size plates.
• Source-space proof comes first: master plate + mask/layer overlay at native coordinates.
• Runtime proof comes second: exact timestamp/frame screenshot from the same harness the user is inspecting.
• Do not claim fixed from intuition. Inspect the generated PNG/crop.

Workflow

1. Identify the master plate and every derived layer/mask involved. For the current scene:

   • assets/scenery/blue_hour_cloudless.webp
   • tools/scenery_art/scenes/blue_hour_waterfront/masks/*.png
   • assets/scenery/*.webp

2. Run the registration probe before editing:

   python3 .claude/skills/scenery-layer-registration/scripts/registration_probe.py \
     --master assets/scenery/blue_hour_cloudless.webp \
     --layer assets/scenery/city_bridge.webp \
     --mask tools/scenery_art/scenes/blue_hour_waterfront/masks/city_bridge.png \
     --crop 430,260,700,440 \
     --jet-time 41.388 \
     --out-dir tmp/scenery_work/registration_probe
   

   Inspect the output overlay/crop. If a real source object is visible in the master but absent from the mask, the mask is incomplete. If the mask covers a non-existing object, the mask is ghosted. If the mask shape is shifted from the source object, the mask is deregistered.

3. Recreate faulty masks from the master, preferably with deterministic image processing. Use OpenCV at build time, not in the app:

   python3 -m venv /tmp/lotti-scenery-opencv
   /tmp/lotti-scenery-opencv/bin/python -m pip install \
     -r tools/scenery_art/requirements.txt
   

   Segment in native 2560x1440 coordinates. Write the mask as a full-frame grayscale PNG. Save a red overlay preview over the master and inspect it.

4. Bake runtime layers from the full-frame masks:

   python3 tools/scenery_art/layer_from_masks.py \
     --master assets/scenery/blue_hour_cloudless.webp \
     --out-dir assets/scenery \
     --preview-dir tmp/scenery_work \
     --layer city_bridge=tools/scenery_art/scenes/blue_hour_waterfront/masks/city_bridge.png \
     --layer yacht=tools/scenery_art/scenes/blue_hour_waterfront/masks/yacht.png \
     --layer foreground=tools/scenery_art/scenes/blue_hour_waterfront/masks/foreground.png
   python3 tools/scenery_art/bake_city_windows.py
   python3 tools/scenery_art/inspect_layers.py
   

   Or just rerun the full blue-hour target (same steps, no OpenCV needed — the base plate is already cloudless, so isolate_clouds.py is not part of it):

   make -C tools/scenery_art blue-hour
   

5. Render the exact problem timestamp, not a nearby approximation. For the dance export harness:

   rm -rf build/character_video_exports/exact_check_frames
   DANCE_EXPORT=1 \
   DANCE_EXPORT_WIDTH=1600 \
   DANCE_EXPORT_HEIGHT=900 \
   DANCE_EXPORT_FPS=60 \
   DANCE_EXPORT_START=41.388 \
   DANCE_EXPORT_DURATION=0.017 \
   DANCE_EXPORT_KEEP_FRAMES=1 \
   DANCE_EXPORT_CAPTIONS=0 \
   DANCE_EXPORT_X264_PRESET=ultrafast \
   DANCE_EXPORT_OUT=build/character_video_exports/exact_check.mp4 \
   fvm flutter test test/features/character/dance_video_export_test.dart
   

   Inspect build/character_video_exports/exact_check_frames/000000.png and a crop around the issue.

6. Only after visual proof, run focused checks:

   fvm flutter test test/features/scenery/scenery_assets_test.dart
   fvm flutter test test/features/scenery/layers/distant_jet_layer_test.dart
   fvm flutter analyze lib/features/scenery test/features/scenery
   

Debugging Rules

• If an object appears in the master but not in the alpha mask, regenerate the mask. Do not add a one-off runtime exception unless the art contract changes.
• If an alpha mask contains objects that do not exist in the master, remove them at the mask-generation source. Do not compensate with opacity or draw order.
• If a moving object should be behind a fixed structure, prove the structure alpha exists at the moving object's source-space footprint before touching z-order.
• If the live app and export differ, use the export/test harness first. Do not run the GUI app just to inspect frames unless the user asks.
• If a helper command fails because cv2 is missing, install/use the documented /tmp/lotti-scenery-opencv venv and rerun the same command.

Helper Script

scripts/registration_probe.py validates layer/mask dimensions and writes:

• full-frame red alpha overlays over the master;
• optional scaled crops for issue inspection;
• optional source-space 747 footprint at a timestamp, matching the current DistantJetLayer constants.

Use it to create evidence before and after changes.