优化版视频人脸马赛克

Long video files are where the plain CPU mosaic node starts to hurt - it chews through a 1080p clip one frame at a time on a single thread. This variant is the pack's answer to that: same Haar-cascade detection, but with a thread pool, frame batching, and a GPU path it tries to use. In the menu it's 优化版视频人脸马赛克 (optimized video face mosaic) under YZZ_Face_Mosaic.

How it works

The core detector is still OpenCV's Haar cascade, but the plumbing is different. With parallel_processing on, frames are read in batches (up to 1000 in memory to avoid blowing up on long clips) and handed to a ThreadPoolExecutor capped at 8 workers - detection for many frames runs concurrently while the writer stays serialized. batch_size (1–16) sets how many frames go through the pipe per batch.

optimization_level is the interesting knob, because it rewrites your other settings:

• fast - forces detection_scale up to at least 1.2 and min_neighbors to at least 3. Fewer detections per frame, faster.
• accurate - clamps detection_scale down to 1.05 and min_neighbors to 8 max. More thorough, slower.
• balanced - leaves your values alone.

So if you set detection_scale to 1.02 and then pick "fast," the node quietly ignores your value. That's worth knowing when you're wondering why faces are getting missed.

About that GPU toggle

use_gpu here attempts OpenCV's CUDA-accelerated CascadeClassifier - cv2.cuda. Reality check: stock pip install opencv-python wheels are not built with CUDA, so hasattr(cv2, 'cuda') is False on almost every ComfyUI install and the node just prints a warning and falls back to CPU. It's a genuine capability, but you'd need a CUDA-enabled OpenCV build to ever see it. Don't buy the node for this feature.

Inputs and outputs

Same cast as the plain node: video_path, mosaic_size, detection_scale, min_neighbors, mosaic_type, output_format, output_quality, optional output_path - plus use_gpu, parallel_processing, batch_size, optimization_level. Outputs are output_video_path and processing_info (JSON with frames, faces, and processing time/fps). Files land in ComfyUI/output/face_mosaic/.

Installing it

Standard pack install: ComfyUI Manager → yzz_face_mosaic, or git clone https://github.com/yzzky/yzz_face_mosaic into custom_nodes + pip install -r requirements.txt, restart. No extra backends needed - this is a pure OpenCV node.

Where it bites

"Optimized" is doing a lot of work. The thread pool helps when the bottleneck is per-frame detection, but video write and OpenCV internals are still serialized, so don't expect a 4× speedup - on many machines the plain node and this one land close. The honest win is optimization_level, which trades detection quality for speed without you fiddling three sliders. If your real problem is accuracy (missed faces), this node doesn't fix that - that's what the InsightFace or MTCNN variants are for.