APNext H3 Stitch Clips (gapless, from the saved files)

After a multi-scene H3 run you've got a folder of per-scene mp4s, each perfect. Then you stitch them and there it is: a tiny silence, a click, a hairline gap at every join. That's not your footage being wrong - it's AAC. Every AAC stream carries ~23 ms of encoder priming and tail padding, and a player only trims that at the start and end of a file. So a plain concatenation - ffmpeg -c copy, an editor's "append" - leaves a 2–14 ms hole at every seam. Your clips are sample-exact; the container isn't. H3StitchClips is the node that puts the seams back together properly.

How it works

It runs once after the last scene and writes <project>_full.mp4 next to the clips. The H.264 packets are copied - their timestamps shifted onto one timeline, no re-encode, no quality loss, no frames held in RAM. Then the audio is handled the clean way: either the original song (wire Load Audio into song) is muxed once over the whole video from 0:00 - the cleanest possible sound - or, with song unplugged, each clip's own audio is decoded, cut to exactly that clip's frame count, concatenated, and encoded once. Either way there's nothing at the joins for a player to trim. The pack verified it on a five-clip run: 773 frames, song at +0.00 ms, no seam silences.

Inputs

• file_paths - the saved clips in order, from H3 Sample + Save's (or Save Clip's) file_path output, one per scene. This is a list, so wire the whole output across.
• suffix (default _full) - appended to the clips' common name: Project-abcd_00001_.mp4 … becomes Project-abcd_full.mp4, in the same folder.
• enabled (on) - off just passes the clips through untouched and writes nothing.
• sync_offset_ms (0) - slides the whole soundtrack against the picture. Negative = the sound comes earlier, positive = later. This is the control to reach for when the lips move a touch before you hear the word - try −40 to −80. (One frame is 41.7 ms, and audio leading the picture is noticed from ~45 ms while audio lagging isn't noticed until ~100 ms, so a small lead reads worse than a small lag.) It belongs here and not on the latent node, because the latent is sample-exact to the picture; a consistent lead is the model anticipating the sound, and the mux is where you trim a constant away without changing what the model heard.

Outputs: file_path (the stitched file) and report (what it did).

The one constraint

The clips have to share resolution and codec - and they do, because they came from one run. Don't try to stitch clips from different projects or with different formats; it'll refuse or quietly do the wrong thing. This node is designed as the last step of a single run, not a general video joiner.

Installing

Part of the dagthomas pack - ComfyUI Manager, search "comfyui_dagthomas", install, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas
pip install -r requirements.txt

No extra dependencies beyond the pack - it uses the ffmpeg machinery ComfyUI already ships. If you've been hand-stitching clips and hunting seam clicks, this is the node that makes the whole pipeline one continuous graph: Sample + Save on the way in, Stitch Clips on the way out, and nothing between them but a finished music video.