Video Watermark Advanced (高级水印)

VideoWatermarkAdvanced is the plain VideoWatermark node with a tracking brain bolted on. Same text/logo stamping, same 22 animation effects, same local OpenCV pipeline - plus the ability to make the watermark chase a person through the frame, dodge busy areas, and process a whole folder of videos in one pass. If you're watermarking clips where the subject moves (which is most of them), a fixed corner stamp is easy to crop out; a watermark that rides along with the subject is a different problem for whoever wants to steal it. That's the pitch, and for the face/body case it mostly works.

How it works

Everything from the base watermark node applies, and on top of that:

• tracking_mode - none, face, body, or object. Here's the honest bit: face and body tracking are implemented with OpenCV's Haar cascades (the classic haarcascade_frontalface_default.xml style classifiers that ship with opencv-python). They're fast and free, and they're also 2010-era tech - frontal faces track great, profile shots and heavily rotated subjects will lose the lock. The object mode is a stub (a TODO in the source), so treat it as "not really implemented." Don't build your workflow on it.
• tracking_position / tracking_offset_x / tracking_offset_y - place the watermark above, below, left, right, or on_target relative to the detected target, with pixel offsets for fine placement (default offset -50 puts it above a face).
• smooth_tracking / smoothing_factor - the node smooths the bounding box between frames so the watermark doesn't jitter; smaller factor = smoother, but laggier.
• enable_anti_occlusion / occlusion_detection_method - try to shift the watermark away from bright/edge/salient regions (brightness, edge, or saliency scoring). Experimental, per the pack's own framing.
• min_target_confidence - 0.1–1.0, default 0.5; raise it to require more confident detections (fewer false locks, more dropouts).
• batch_mode / batch_video_folder / batch_output_folder - process every video in a folder. Great for a day of generated clips; mind the disk space.
• text_effect - none or curved, for arc-shaped text. It's the little flourish the basic node lacks.

Outputs: video (VIDEO) and info (STRING).

Install

Same pack install - ComfyUI Manager → "ComfyUI_Sora", or clone + pip install -r requirements.txt. No API key needed; Haar cascade files come bundled with opencv-python, so there's nothing extra to download.

Common issues

• Tracking loses the face mid-clip - expected with Haar cascades on side/rotated faces. Keep the subject front-facing, or fall back to fixed position for anything stylized.
• object mode does nothing - it's a stub; use face/body or the fixed multi-position presets instead.
• Watermark flickers between positions - that's enable_multi_position fighting tracking_mode; the tooltip even warns they conflict. Pick one.
• Batch runs forever / fills the disk - batch mode re-encodes every video; check batch_output_folder has room and that you actually want all of them.

The honest verdict: for a fixed or animated watermark, the base node is plenty. Use this one when your clip has a clear frontal subject worth pinning a watermark to - it's a neat trick when it locks, and a "turn it off" when it doesn't.