Nodes/minimaxH3DYTsc6.0/minimaxH3DYTsc6.0 VideoCompare
ComfyUI Node

minimaxH3DYTsc6.0 VideoCompare

The node that finally gets your H3 audio into the MP4

By 792877530-star·Created 26 days ago·Updated 26 days ago· 2
minimaxH3DYTsc6.0 VideoCompare
  • p2_images
  • p1_images
  • video_images
  • video_audio
  • images
◄video_fps24.00►

What it's for

minimaxH3DYTsc6.0 VideoCompare (class MinimaxH3DYTScVideoCompare) is the last node in this pack's chain, and it's the one that turns a pile of frames plus a generated soundtrack into a file you can actually send someone.

It hooks out of the minimaxH3DYTSc6.0 Director node. The Director exposes two comparison outputs - compare_p1 (the first sampling pass) and compare_p2 (the refined, final pass) - and VideoCompare is where you watch them arrive segment by segment and save whichever you prefer. On a run with the second pass disabled, both are the same frames and this is simply your viewer and exporter.

Fun fact about H3 that makes this node matter more than it looks: this is a model whose entire pitch is native stereo audio generated jointly with the picture. Getting that audio into the container is not a detail you want to hand-roll.

Inputs and outputs

p2_images is the only required input, and you connect it to the Director's compare_p2. The optional side is p1_images (the first pass, for the before/after pane), plus video_images, video_audio and video_fps - wire those to the Director's images, audio and fps outputs, and the node can write a finished MP4 with the soundtrack muxed in.

The single output images just passes p2_images through, so it's safe to park at the end of a graph or chain into something else.

What happens when you hit save

The node registers itself with the backend under its own node id and hands the frontend panel the per-segment frames; the panel draws them as they finish, with separate save buttons for the first and second pass. Those buttons hit the pack's own HTTP route, which pulls the frames for that pass, concatenates the Director's per-segment audio waveforms into one track and encodes with the ffmpeg binary that imageio-ffmpeg ships - H.264 and AAC, CRF 18, written to output/minimax_h3_compare/. No system ffmpeg required, which is the one thing this pack does get right about dependencies.

Frame rate comes from the Director's fps output, falling back to the node's video_fps widget (24 by default). If the audio path comes up empty the file is still written, just silent - you get a video with no track rather than an error.

Install

It's part of the same pack, so there's nothing extra to fetch:

cd ComfyUI/custom_nodes
git clone https://github.com/792877530-star/ComfyUI-MinimaxH3DYTsc

Restart ComfyUI afterwards - the Python side (the save route, the ffmpeg encode) and the JS side (the panel) both load at startup. Manager users: search minimaxH3DYTsc6.0. The pack's requirements - opencv-python-headless, imageio-ffmpeg, scenedetect, comfy-kitchen>=0.2.34 - install with it, and imageio-ffmpeg is the one this node actually needs.

Gotchas

  • Nothing renders in the node body. The panel is a JS extension bound to the exact class name, and the pack's own source notes that an earlier build shipped it under a mismatched name so the panel never attached and an older plugin's script drew the UI instead. Update the pack, then hard-refresh the browser. If you have an older build of this plugin (the "v141" line) sitting in custom_nodes next to this one, remove it - the two families share frontend hooks.
  • "no finished video data yet" when you hit save: you either haven't run anything or you're saving a pass that never produced frames. Run the Director first, then save.
  • Silent MP4. The soundtrack comes from the Director's audio output, which itself depends on the timeline's audio mode - a source-audio mode with no usable source audio gives you nothing to mux. Check the Director's audio settings before blaming this node.
  • It's a viewer, not a diff tool. There's no blend or slider across the whole clip; if you want a proper A/B of single frames, the ecosystem's muscle memory for that is rgthree's Image Comparer or core's Image Compare node. This one is for watching a render land segment by segment and exporting it.
  • Many error strings in this pack are in Chinese. They're specific and accurate - 该路暂无成品视频数据 means "that pass has no frames" - but they're not searchable.
CategoryMiniMaxH3

Inputs (5)

NameTypeDefaultDescription
p2_imagesIMAGE连接导演台的 compare_p2 输出(每段二采/最终成片)。
p1_imagesoptIMAGE连接导演台的 compare_p1 输出(每段一采成片)。
video_imagesoptIMAGE连接导演台的 images 输出(最终成片帧,用于保存成品视频)。
video_audiooptAUDIO连接导演台的 audio 输出(最终成片音轨,保存成品视频时混入)。
video_fpsoptFLOAT24.00连接导演台的 fps 输出(最终成片帧率)。

Outputs (1)

NameTypeDescription
imagesIMAGE—