Memory optimization + fixes

[Elricfae]

Summary

These changes focus on performance optimization and code quality improvements, with a particular emphasis on VRAM management, CUDA efficiency, and code formatting consistency. The modifications span multiple processor modules and optimize memory allocation patterns throughout the face processing pipeline.

Key Changes:

🚀 Performance Optimizations:

• VRAM/PCIe Efficiency: Replaced CPU-allocated tensors with direct GPU allocation using torch.ones(), torch.zeros(), and torch.full() to avoid unnecessary PCIe transfers
• Memory Reuse in DDIM Loop: Pre-allocated buffers that are reused in-place (.fill_() method) instead of creating new tensors on every step, eliminating VRAM fragmentation
• Lazy Kernel Initialization: Converted convolution kernels to lazy-initialized properties, ensuring thread-safe CUDA allocation without pre-allocation overhead
• Video Buffer Management: Optimized frame buffer sizing from preroll_target * 4 to preroll_target + (num_threads * 2) with RAM safety net checks to prevent OOM on high-resolution videos

🧹 Code Quality:

• Standardized multi-line formatting for io_binding.bind_output() calls across all face detection/swapping models (RetinaFace, SCRFD, YOLOFace, YuNet, CSCS, etc.)
• Removed redundant .clone() operations where references suffice (e.g., mask assignments, tensor assignments)
• Unified tensor operations using in-place methods and cleaner chaining patterns
• Fixed redundant CUDA stream sync comments and removed dead code

** 🐛 Batch Video Processing Fixes

• Face Mix-up Prevention: Added force_recognition_in_batch flag to disable fast-path optimization during batch processing, ensuring true identity verification on each video
• State Reset: Detector and Temporal EMA states now purged between batch videos

🔧 Minor Fixes:

• Adjusted eye ratio minimum threshold from 0.10 to 0.08 for better eyelid blending
• Suppressed non-critical ONNX warnings by consistently setting log_severity_level = 3
• Streamlined ONNX session initialization logic
• Input Face Loader Robustness: Added try-except wrapping in the loader worker to gracefully skip corrupted images instead of crashing
• Input Face UI Batching: Implemented batched UI updates with a Python deque + timer system to prevent main thread freezing during large face loads

📊 Expected Impact

• VRAM: Noticeably reduced fragmentation and peak memory usage, especially during multi-iteration operations
• Throughput: GPU kernels execute with less synchronization overhead
• Reliability: Batch processing now handles mixed-quality videos without face identity confusion
• UX: Input face loading no longer freezes the UI on large datasets