Video encoding separation to new file

[Elricfae]

Overview

This PR refactors the FFmpeg encoding pipeline by extracting subprocess management into a new dedicated module (video_encoding.py), reducing code duplication by ~500 lines and improving testability and maintainability.

🎯 Key Changes

1. New Module: FFmpegEncoder Class (video_utils/video_encoding.py)

• Encapsulates FFmpeg subprocess lifecycle: process creation, frame writing, and graceful shutdown
• Implements robust timeout handling with 3-step escalation (graceful close → SIGTERM → SIGKILL)
• Handles source video probing with caching to prevent redundant I/O calls
• Manages quality adaptation logic (CQ/CRF calculation from source metrics)
• Provides public interface: start_process(), write_frame(), close_process(), is_running()

2. New Module: FFmpegPostProcessor Class (Stateless utilities)

• extract_audio_segments() – Extracts per-segment audio with validation
• concatenate_audio_segments() – Merges audio using FFmpeg concat demuxer
• write_video_only_output() – Fallback video-only muxing when audio processing fails
• concatenate_segments_video_only() – Segment-mode video concatenation fallback
• validate_audio_file() – Validates audio integrity before processing

3. Refactored video_processor.py

• Replaced inline FFmpeg subprocess calls with self.encoder.start_process()
• Replaced direct stdin writes with self.encoder.write_frame() and self.encoder.is_running()
• Simplified cleanup: self.encoder.close_process() handles all subprocess termination
• Removed obsolete create_ffmpeg_subprocess() method (functionality moved to FFmpegEncoder)
• Removed redundant audio extraction methods (now in FFmpegPostProcessor)

4. Enhanced Face Expression Processing (frame_edits.py, faceutil.py)

• Gaze Stabilization Pipeline: New "Camera Gaze Lock" feature forces eyes to look at camera with per-frame adjustment
• Eye Aspect Ratio (EAR) Safeguards: Profile occlusion detection, symmetric blink harmonization, anti-bulging eyelid compensation
• Mouth Aspect Ratio (MAR) Safeguards: Division-by-zero prevention, biological plausibility clamping
• Split-Eye Calculation: Enable asymmetric winking via independent left/right eye retargeting
• Structural Lip Decoupling: Separate relative (X-axis shape) and absolute (Y/Z-axis jaw) motion with blending
• Removed redundant GPU stream synchronization after processing

5. UI/UX Improvements

• Added "Camera Gaze Lock" controls to advanced face expression settings
• Added keyboard shortcut X for Face Compare mode
• Fixed keyboard navigation to use step-through pipeline (prevents flickering in Compare/Mask modes)
• Improved slider preview behavior: suppresses raw frame flash when stepping with special modes active
• Fixed arrow key navigation to ensure single-frame stepping instead of slider defaults

6. TensorRT Cache Cleanup (models_processor.py)

• New _clean_tensorrt_cache() method safely removes corrupted engine/timing files
• Prevents "cache poisoning" failures on retry after timeout
• Handles locked/missing files gracefully

✅ Benefits

• Testability: FFmpeg logic is now isolated and unit-testable
• Maintainability: Centralized encoding/audio handling vs. scattered throughout codebase
• Extensibility: Easy to add new post-processors (subtitles, metadata, etc.)
• Robustness: Explicit lifecycle management prevents zombie processes and timeout cascades
• Performance: Source metrics caching eliminates redundant ffprobe calls
• User Experience: Smoother navigation with reduced UI flashing