### 3D Pipeline Improvements – v3.3

- **Black Bar Handling**
  - Detects and crops letterbox bars before depth estimation.
  - Re-adds bars post-processing with neutral depth to prevent stereo artifacts.

- **Depth Stability & Quality**
  - Added temporal depth filtering (EMA) for smoother frame-to-frame depth.
  - Percentile-based normalization for consistent depth range across scenes.
  - Midtone shaping to enhance perceived depth and separation.
  - Optional curvature enhancement for more natural object roundness.

- **Stereo & Parallax Control**
  - Dynamic IPD scaling based on scene depth variance.
  - Shift smoothing for foreground, midground, and background layers.
  - Edge-aware masking and feathering to reduce tearing/ghosting.
  - Subject-tracked zero-parallax plane with eased floating window adjustments.
  - Optional dynamic convergence bias tied to subject position.

- **Image Quality**
  - GPU-accelerated depth-of-field with multi-level Gaussian blending.
  - Brightness-preserving sharpening with anti-clipping.

- **Framing & Output**
  - Per-frame optional black bar cropping in the stereo stage.
  - Aspect-ratio-safe resizing and padding for per-eye frame alignment.
  - Multiple stereo formats supported (Full-SBS, Half-SBS, VR, anaglyph, interlaced).

- **Encoding & Stability**
  - FFmpeg pipeline over stdin with correct CRF/CQ handling.
  - Graceful cleanup and resource release on completion or cancel.
  - Detailed logging for crop, skip, and encoder decisions.

### Depth Pipeline Improvements – v3.3

- **Letterbox Detection & Cropping**
  - Added `ignore_letterbox_bars` option to automatically detect top/bottom black bars.
  - Crops bars before sending frames to the depth model for cleaner predictions.
  - Stores bar metadata (top/bottom size, original resolution) in a `.letterbox.json` sidecar file.

- **Preserve Original Resolution**
  - After depth prediction, resizes the depth map back to the cropped region size.
  - Re-adds original bars with a neutral depth value to avoid 3D distortion.
  - Ensures depth output matches original width/height exactly.

- **Crash Prevention & Safety Checks**
  - Resets bar values to zero if detections are invalid (too large or empty frames).
  - Handles fade-ins and all-black first frames without breaking processing.

- **Grayscale Conversion Refactor**
  - Unified `convert_depth_to_grayscale()` into a single reusable function.
  - Works with `PIL.Image`, `torch.Tensor`, and `numpy.ndarray`.
  - Includes NaN checks, shape handling, and safe fallbacks for bad frames.

- **Output Consistency**
  - Standardized grayscale depth conversion across all depth output paths.
  - Maintains uniform depth range handling for downstream 3D processing.

Version Tag Change