视频网格拼接

The fastest way to evaluate a motion-transfer result is to put the source pose video and the generated video side by side and watch them together. VideoGridCompose is exactly that: it takes up to four video file paths, arranges them in a grid via ffmpeg, and writes a single composited file to ComfyUI's output folder. Because it streams through ffmpeg's filter graph instead of decoding every frame into the graph, it handles long clips without eating VRAM or RAM.

The classic use here is debugging WanAnimate and SCAIL-2 runs: left pane the driving video, right pane the generation, and you can literally see whether the pose was followed, whether the subject drifts, and where the chunk seams are. It's also handy for contact sheets of multiple generations to pick the winner.

The inputs that matter

• video_paths - one path per line in a multiline box, up to 4 (duplicates and blank lines are dropped). Paths can be relative to the ComfyUI root or absolute.
• layout - 水平排列 (horizontal strip), 垂直排列 (vertical strip), 田字格2×2 (2×2 grid), 左一右N (one left, N right), 上一+下N (one top, N below).
• audio_source - which input's audio survives to the output (or none). Nice when you want to compare against the original sound.
• match_method - how mismatched resolutions are reconciled: 自适应(填满) stretches to fill tiles, 自适应(保持比例) fits without distortion, 统一到最大/统一到最小 scale everything to the biggest or smallest input.
• fps (default 25) - output frame rate; 0 picks the highest input rate.
• crf (default 23) - encode quality, lower is better (the tooltip suggests 18–28).
• duration_mode - which source sets the timeline: 自动(最短)/自动(最长)/视频1..4. This matters because mismatched clips otherwise get cut at the first short one. It skips execution if the chosen reference is a still image rather than a video.
• output_path_prefix (default video/grid_output) - where the result lands, relative to ComfyUI's output directory.

The single output is video_path - a string pointing at the composited file.

Where it fits

This is the terminal node of the pack's pose and animation workflows (the README's workflow section calls it after CrossFadeImages merges chunk outputs). Keep crf reasonable - 23 is fine for review, drop toward 18 if you're going to upload the comparison. If tiles come out letterboxed in a weird way, check match_method: for mixed resolutions, 保持比例 is usually the honest view.

Install

Part of ComfyUI-CustomNodeKit. ComfyUI Manager → search "ComfyUI-CustomNodeKit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

Restart ComfyUI. ffmpeg comes via the pack's imageio-ffmpeg dependency, so no system install needed.