H3 Stitch Frames (internal)

The classic way to compare two AI videos is to tab back and forth and hope your eyes remember. The pack's Compare & Stitch page does it properly - pick 2–4 clips from the Library or the outputs strip, compare them playing side-by-side in sync, or export them as a single stitched clip. H3 Stitch Frames (internal) is the node that does the actual stitching, and it's deliberately boring: its only job is to take up to four IMAGE batches and lay them out on one canvas.

Inputs:

• image_1 (IMAGE, required), plus image_2, image_3, image_4 (all optional) - the clips to combine.
• padding (INT, default 8, range 0–256) - the gutter width between cells.
• pad_frames (enum, freeze/black) - what to do with shorter clips.
• columns (INT, default 0, range 0–4) - 0 means auto (roughly a square grid); 1–4 forces that many columns.
• background (STRING, default #000000) - the canvas color behind everything.

Output: a single images (IMAGE) batch.

The mechanism is where the quality lives. Cell size is the max width and height across all clips, so every source is contain-fit into the same-sized cell - letterboxed, centered, never cropped, with the gutter and background color filled in around them. Shorter clips are padded to the longest: freeze repeats the last frame (which is what you want for a comparison where one take is longer), black pads with zeros. Frame matching itself is done upstream by the VideoHelperSuite loaders - a deterministic 24 fps plan that computes skip_first_frames and frame_load_cap per clip so the loaded windows line up by construction. That's why the stitched comparison is actually in sync instead of "close enough." The combined clip's audio comes from clip 1, and export is h264 or h265 mp4.

You won't add this node to a canvas by hand. When you hit Compare or Stitch in the UI, the pack's backend stages your picked outputs into the input folder, probes their real frame counts, builds a VHS_LoadVideo → H3StitchFrames → VHS_VideoCombine graph, and queues it through the normal prompt path. The one dependency it leans on is ComfyUI VideoHelperSuite (for those loaders and the combine) - and the pack already lists that as part of its stack, so if you've used the Library or preview-without-saving you've got it.

The design choice worth appreciating: the stitch node itself has zero ComfyUI imports and could be lifted straight into any graph - it's just padding and grid geometry, which is exactly the kind of plumbing the ecosystem (and the KB) keeps rediscovering. If you ever need a dependency-free side-by-side stitcher outside this pack, steal the idea.