Save Clip

Save Clip is the video half of Kinburg-Nodes' song suite - the sibling of Save Song, built to sit at the end of the pack's Siren chain, where ACE-Step-class audio models turn lyrics and a section plan into an actual track. You feed it a picture (or a whole batch of pictures) and that audio, and it writes an h264+AAC mp4 whose cuts land where the music changes. One image and a three-minute song gets you a single still held over the whole track; a batch of images gets you a slideshow whose timing follows the song.

It's one of those nodes that sounds trivial until you price out the obvious alternative. In plain ComfyUI, "image batch plus audio" means Create Video + Save Video, and those encode one frame per batch item. A three-minute song at 30 fps is a 5,400-frame batch - roughly 17 GB of tensor for a 1024×1024 picture before the encoder has even seen anything. Save Clip never builds that batch. It keeps at most a couple of prepared slides alive and feeds the encoder frame by frame through PyAV (av, which already ships inside ComfyUI), so its memory footprint is set by the frame size, never by the song's length.

How the timing works

Here's the part worth understanding before you touch it: the audio's real length is the video's length, and the plan only supplies proportions. A Siren plan is written in bars ("Chorus | Nina | 8 bars"), and bars are only seconds once you know a tempo that lives on some other node - the moment the two disagree, your pictures slide off the music. So the plan is read for its ratios and scaled so the last slide ends on the last sample. No BPM to wire, no drift, and if the plan came out 5% short, the stretch factor is printed in the report instead of silently wrong. Leave plan empty and it just splits the song evenly.

The inputs that actually matter:

• image and audio are the required pair. Nothing resamples in time - the batch is the list of pictures, not the video's frames. Four images over a three-minute song is four shots, not four frames.
• layout - "one slide per section" (the default), "by section label" (every row called Chorus gets the same slide, so the chorus shot comes back like it does in a cut music video), or "even" (ignore the plan).
• crossfade (default 0.5s, centred on the cut) and ken_burns (off by default) are the only things that actually move. Ken Burns makes every frame unique, which grows the file and encode time several times over - with it on, raise fps to 24+ or the motion steps instead of flows.
• lyrics with their [Verse 1]-style markers writes an .srt next to the mp4, timed to each section - but only when a Siren plan is wired, because the lyrics' markers and the plan's labels are matched on the same words.
• tags takes a SONG_TAGS bundle (the same node Save Song eats) and writes it into the mp4's own metadata.

You get three outputs: video (a VIDEO object on newer ComfyUI builds, so the clip can chain into ComfyUI's own video nodes; on older builds the file is simply on disk), path (where it landed), and report (what happened, section by section).

Installing it

It's one of 98 nodes in the Kinburg-Nodes pack, so you install the whole thing:

cd ComfyUI/custom_nodes
git clone https://github.com/Kinburg/Kinburg-Nodes

or search "Kinburg-Nodes" in ComfyUI Manager. One gotcha: the pack's install.py (which Manager runs automatically) installs the CUDA build of llama-cpp-python for its LLM nodes - you don't need that for Save Clip, but it's part of a normal install, so don't panic at the wait. Save Clip's own needs - av, Pillow, numpy - are already in ComfyUI.

Where people get burned

The batch-isn't-frames confusion is the big one. Ken Burns left on at 12 fps looks steppy; crank the fps. A .srt that doesn't appear almost always means the lyrics' markers and the plan's labels don't literally match - wire both from the same Siren node. An empty audio input fails loudly with "there is nothing to time the slides to." And odd source dimensions get rounded down to even sides, because yuv420p demands it. The timing logic is the strongest part of this node - trust the audio, not your plan's bar count.