MiniMax H3 Video Merge (Concat)

So you used MiniMax H3 Video Extender to generate clip 2 as a continuation of clip 1, decoded both, and now you have two frame sequences and two audio tracks that are supposed to be one video. A naive concat gets you the telltale seam: a one-frame brightness pop where clip 2's exposure doesn't match clip 1's tail, a hesitation on clip 2's first generated frame, and - if your audio came from a file on one side and a latent decode on the other - sample rates that don't even agree. MiniMax H3 Video Merge (Concat) is the "join them cleanly" step in this pack: it takes the two clips, fixes the seam, and hands you one images stream and one audio stream ready for your video encoder.

What it does

images_b is the current, newly rendered clip (required). images_a is the previous clip (optional). The naming trips everyone once: a = old, b = new. If images_a is empty the node passes images_b straight through, so it's safe to leave sitting in a single-clip workflow. When both sides are present, it does four things in order:

1. Resolution-aligns. If the two clips differ in size, images_a is upscaled with Lanczos to match images_b's canvas - so the new clip's resolution is the one that wins.
2. Color-matches images_b to images_a. The default color_match = "seam_fade" matches luminance and chrominance at the exact boundary and cosine-fades back to the clip's native colors over ~24 frames - that's what kills the brightness flash. match_mean_std remaps B's whole tone to A's when the clips have drifted beyond the seam; disabled keeps native colors.
3. Seam-smooths motion. seam_smooth (default 1) runs Farneback optical flow across the cut: it nudges clip B's first frame - the one diffusion models hesitate on - forward along its own motion and balances velocity at the boundary, without dropping frames, so your total count and audio timeline stay intact. Radius 2–3 smooths a wider window.
4. Concatenates audio. If both audio_a and audio_b are present, it resamples to a common sample rate via torchaudio, pads any mono↔stereo mismatch, and joins them.

Outputs are images and audio, which you feed into whatever node actually saves the file (VHS VideoCombine and friends). Nothing in this node encodes video - it's purely the stitch.

The inputs that matter

• seam_smooth - 0 turns optical flow off entirely; 1 smooths just the seam frames; 2–3 widen the transition. This is the one you'll fiddle with when motion feels sticky at the cut.
• color_match - leave on seam_fade unless you can see a whole-clip tone mismatch that seam matching doesn't cure.
• blend_frames - default 0, and keep it there while seam_smooth ≥ 1. Above 0 it switches to a true crossfade: the last k frames of A and first k of B get blended and the raw tails/heads dropped. It's an alternative to seam smoothing, not a stacking bonus - running both double-processes the boundary. fps only matters here, converting the frame crossfade into the matching audio crossfade duration.

Where it sits in the chain

This node is the back half of a two-node pipeline. The Extender pins your previous clip's tail into the new sample and outputs a trim_frames count; you decode, run MiniMax H3 Extension Trim to drop those overlap frames, then feed the result in as images_b with the previous clip as images_a. Went the pure-latent route? MiniMax H3 Load AV Latent hands you both images_a and audio_a with no VAE round-trip - which is why the pack's demo workflows lean on it for multi-step chains.

Install

Merge ships in the same pack as the Extender, so the install is identical: ComfyUI Manager → search MiniMax H3 Extender → Install → restart, or

cd ComfyUI/custom_nodes
git clone https://github.com/pmhaidn/ComfyUI-Minimax-H3-Extender.git
pip install -r requirements.txt

That requirements file is a single line - opencv-python - and it genuinely matters here.

Common issues

• Seam smoothing looks like a cheap crossfade. The optical flow path needs OpenCV; if opencv-python isn't installed, the node logs a warning and silently falls back to a plain linear blend. If Manager didn't pull deps, install it manually and restart - the difference is real.
• Still seeing a color pop at the seam. Check both clips decoded through the same VAE pipeline, then try color_match = "match_mean_std" if the mismatch is whole-clip rather than boundary-only.
• Audio is there but sounds slightly off. Merged audio is only as good as what you feed it. If clip A's audio came from an MP4 and clip B's from a latent decode, sample rates get aligned but the recordings won't be identical - the pure-latent continuation path upstream is what buys true continuity.
• Final video came out at the wrong resolution. Remember images_a is upscaled to images_b - generate your continuation at the resolution you actually want the final cut to be, or the old clip gets silently resized to match.