Video Save Plus (Custom Path) 🎬

ComfyUI's default video output has a hole: it saves the pixels and, depending on which node you grabbed, quietly drops the soundtrack. That used to be fine - most video models were silent. Then MiniMax H3 landed with native stereo audio built in (the first serious open answer to Veo's audio gap), and "where's the sound?" became the most annoying question in the room. VideoSavePlus is the answer. It's a video output node that writes an MP4 with H.264/H.265 plus AAC audio to any folder on any drive, aimed squarely at the new class of models that hand you a VIDEO tensor with audio already attached.

Why v2 exists

This pack is a rewrite, and the changelog is the story. The old node, VideoCombineCustomPath, wrote audio through torchaudio.save, which silently drops the track on newer torchaudio versions - you'd get a pristine silent MP4 and never know. v2 pipes raw float32 PCM and raw RGB frames straight into ffmpeg over pipes: no torchaudio, no temp PNG sequence, and if an audio input can't be processed it raises a clear error instead of quietly succeeding. "Fails loudly" is the right design for a save node, because a silent "successful" render is worse than a red one. One-time cost: old workflows have to be re-wired to the new node name once.

What to plug in

• video (VIDEO) - the output of native-audio models like MiniMax. Frames, audio, and frame rate are all pulled from it.
• images (IMAGE) - the fallback for classic image-batch workflows; frame_rate only applies here, since a VIDEO input brings its own.
• audio (AUDIO) - overrides the track inside a VIDEO input, handy for dubbing a silent generation.
• filename_prefix - supports subfolders (scene1/take) and the placeholders %date%, %time%, %seed%, %date:yyyy-MM-dd hh-mm-ss%. A 5-digit counter is always appended.
• custom_output_path - any folder on any drive; created if missing, empty falls back to ComfyUI's output dir.
• crf (default 19) - the main quality knob: 17–19 is visually lossless, 0 is lossless, 28 is small.

The output is a single filepath string with the full path of the saved MP4 - wire it to a Text or a "Show Text" node if you want to see it in the UI.

The stuff that makes it a keeper

Beyond the video itself, each run writes a workflow PNG (drag it back onto the canvas to rebuild the graph - the same workflow-in-the-file convention ComfyUI's own SaveImage uses), an optional name.txt with your prompt, and optionally frames as first/last/all in jpg or png. That's the whole training pair in one folder, which is exactly what the Save Training File button is for: it copies the mp4 + png + txt + frames to copy_to_folder under identical names. There are also buttons to reveal the file in your file manager, play it with sound, grab the last frame losslessly, and delete the whole set - and the server endpoints only ever touch files this node saved in the current session, not arbitrary paths.

Install

Via ComfyUI Manager, search for Video Save Plus, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Slartibart23/ComfyUI-CustomVideoSave

Restart ComfyUI and hit Ctrl+F5 in the browser once. The only hard dependency is ffmpeg on your PATH (or pip install imageio-ffmpeg as a fallback - the node checks both); requirements.txt is just pillow, so no torchaudio install to fight.

Gotchas

• h265 is ~40% smaller at the same CRF but many players and browsers refuse it - h264 plays everywhere, and yuv420p is the compatibility-safe pixel format. yuv444p looks sharper on text/edges but is a browser no-go.
• Odd frame dimensions get auto-scaled to even numbers for yuv420p - a classic ffmpeg footgun this node just handles for you.
• If ffmpeg is missing, you get a loud error, not a silent file. That's the feature.