PJ-视频与标签保存 (Batch Video Saver)

Most "save video" nodes drop a file and call it a day. If you're building a dataset or organizing generations, that's the wrong half of the job - you want the clip and its description to land together, paired by filename, the way every image trainer pairs image.png with image.txt. BatchVideoSaver (pongjoo/ComfyUI-PJLatent, menu "PJ/批量视频") writes both in one pass: the frame batch becomes an mp4 or avi, and the tag text you feed it becomes a same-basename .txt sitting right next to it. That sidecar convention is exactly the training-set pattern from image-io-metadata.md, applied to clips.

It's an output node - no outputs, nothing to wire onward. You feed it frames and text, it writes files.

The inputs that matter:

• 图像流 (images) - the IMAGE batch. Typically frames you extracted from a video and processed (upscaled, interpolated) in the graph above.
• 标签文本 (tag_text) - force-input, so wire it rather than typing. In a batch pipeline this comes from whatever described your content: a tagger or VLM node on the first frame, or a text node fed by the batch path provider's filename.
• 附加文本 (extra_text) + 附加位置 - an optional note appended at the start or end of the tag file. Handy for generated with <workflow> stamps.
• 帧率 (fps) - defaults to 16. That's the video-latent convention in this pack (seconds × 16 + 1 frames), not a universal truth; set it to your source's real rate.
• 视频格式 - mp4 or avi.
• 文件名前缀 (prefix), 自定义保存路径, 自定义文件名, 时间戳 (None/Second/Millisecond) - the naming toolbox. 移除自定义后缀 (remove_ext) strips an extension from a custom filename so your .mp4/.txt pair stays consistent.

No filename given? It falls back to ComfyUI's standard counter behavior (video_00001_<timestamp>.mp4).

How it actually writes

Straight through OpenCV's cv2.VideoWriter: mp4 uses the mp4v fourcc, avi uses XVID. Frames go RGB→BGR; if you feed RGBA, alpha is silently dropped. The .txt is written with UTF-8 encoding. If you connect no images at all it still writes the tag file alone - a "text mode" the author built in, not a bug.

Two honest caveats. First, mp4v is MPEG-4 Part 2, not H.264 - it plays everywhere but the files are chunkier than a proper encode, and there's no audio track (there's nothing to write; you fed it frames). If you're producing a portfolio master, re-encode with ffmpeg afterward. Second, the OpenCV dependency: ComfyUI's own requirements pull in opencv-python-headless, so cv2 is normally already there - but if you run a trimmed environment, you'll see No module named 'cv2' the first time you hit Run. pip install it into the same environment as ComfyUI.

The batch recipe

Where this shines is wired to its sibling BatchVideoPathProvider: provider hands out one video path + its clean basename per queue run → load and process the frames → BatchVideoSaver gets the basename as 自定义文件名 and the tagger's output as 标签文本. Every run of the queue drops out one finished clip with its caption, named after the source. That's the workflow people end up hand-assembling with string nodes; the pack just pre-built it.

Install

Part of ComfyUI-PJLatent: ComfyUI Manager → search "ComfyUI-PJLatent", or

cd ComfyUI/custom_nodes/
git clone https://github.com/pongjoo/ComfyUI-PJLatent.git

Restart ComfyUI. No requirements.txt, no models to download. The node shows up as PJ-视频与标签保存 (Batch Video Saver).