Highlight · 高光片段提取

The Highlight node answers a very specific question: where in this video are the moments people would actually keep? It watches for emotional peaks and key actions, then hands you timestamps and scores as JSON - no manual scrubbing through a two-hour recording to find the one good take. It's tuned for exactly two kinds of content, which is worth knowing before you feed it anything else.

The one input that matters

• model - Miniseries (short drama) or Game (casual game footage). That's the entire tuning surface, and it's a meaningful choice: Miniseries hunts narrative beats - emotional peaks, plot-relevant action - while Game is tuned for gameplay moments. Feed a drama into Game and you'll get highlights, just possibly not the ones a story editor would pick. Match the model to the content.

Then the standard trio: video, poll_interval_seconds (default 10) and max_poll_attempts (default 720).

What comes out

Two STRING outputs:

• highlights_json - the highlight metadata: timestamps, scores, and the reasoning-ish detail the cloud model attaches to each peak.
• raw_json - the complete result, with API keys and signed URLs redacted by the pack's redaction pass.

Neither is a video. This node produces metadata, so the job isn't done on the canvas - you pair it with a downstream node or a small script that actually cuts the clip at those timestamps, or feed the JSON into a visualization node if you just want a timeline view. Think "automated trailer outline," not "rendered trailer."

How it works

Standard cloud choreography: the video materializes to a temp file, mediakit-cli submits an async analysis task to your MediaKit account, the node polls on your interval, then downloads the result payload and returns it redacted as two JSON strings. Like every node in this pack, it bills your account by usage and sends the clip to Volcengine - there's no local model, no weights, no GPU load.

Install and the cloud setup

Install via ComfyUI Manager (search "MediaKit Toolkit" / mediakit-toolkit), full restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/fudanll-sys/comfyui-mediakit-toolkit.git

No Python dependencies and no model downloads - the pack wraps Volcengine's CLI. Before any cloud node works you need a Volcengine account with AI MediaKit activated, an API key, and the CLI installed and authed in the environment that starts ComfyUI:

npx @volcengine/mediakit-cli install -y
mediakit-cli init --mode cloud-first --api-key "YOUR_MEDIAKIT_API_KEY" --credential-store config --yes

Where it fits

If you batch-convert game VODs or short dramas and want an auto-cut candidate list, this is a genuinely useful timesaver - it's the kind of analysis that would otherwise cost you an afternoon of scrubbing. The honest limits: "highlight" is a judgment call, so treat the scores as a first pass, not gospel; the output is JSON you have to wire up yourself; and on content that isn't drama or gameplay, results will be mediocre. Start with a short clip to see whether the model's taste matches yours before you run it across a season.