YouTube Trend Topics

YoutubeTrendTopics is the ideation stage of the Story JSON Nodes pack (stepan-bogatorjov/comfy_scenes_json_node), and it answers the question the rest of the pack is built around: what story should we generate next? It calls the official YouTube Data API v3 - no scraping, no downloads - pulls recently published videos for your query and region, and returns a ranked list of topic candidates scored for how well they'd work as AI short-video storytelling.

If the pack is a YouTube-shorts factory, this is the market research department. The default query, "what happens inside", is not random: it's a curiosity-stem title pattern the ranking explicitly rewards. The node fetches video stats (views, likes, comments, duration), computes views-per-hour and engagement, and multiplies by a relevance boost derived from the title - patterns like "what happens if", "why you should never", "the hidden truth", "survival", "explained" all add points. Then it filters the noise: music videos, reactions, livestreams, gameplay, podcasts get dropped or heavily penalized. The output is JSON you can feed straight into the story generator.

How it works

Two API calls: search for video IDs in the window, then videos.list for details. Each result is scored (views_per_hour × 0.7 + engagement × 0.2 + recency × 0.1, times the title boost), then filtered by min_views, exclude/include keyword lists, a Shorts preference (clips ≤90s get a boost, >600s get halved), and a hard language re-check via the lingua-language-detector package - because YouTube's relevanceLanguage is only a hint. It never crashes the graph: problems come back as an {"error": ...} JSON plus an ERROR: text line.

The inputs that matter

• api_key - a YouTube Data API v3 key. Leave empty to use a YOUTUBE_API_KEY env var or a .env file (there's an .env.example in the repo).
• query - the search stem ("what happens inside" by default).
• region_code - ISO code, default US.
• max_results - up to 50.
• published_after_days - how far back to look (default 7).
• order - viewCount, date, or relevance.
• video_duration / safe_search / language - the usual YouTube filters; language defaults to en.
• min_views, exclude_keywords, include_keywords, prefer_shorts, output_format - the ranking and output dials.
• Outputs: topics_json (the full ranked records) and topics_text (a plain "1. title" list). In text mode both outputs mirror the text.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/stepan-bogatorjov/comfy_scenes_json_node

The pack's own dependencies (requests, lingua-language-detector) install with it via Manager. Then get a free YouTube Data API v3 key from the Google Cloud console (APIs & Services → Credentials → Create credentials → API key) - the repo's .env.example is there for exactly this.

Troubleshooting

The key resolution order is input → env var → .env, and the errors are unusually good about telling you which one you skipped - plus, in one of the nicest touches in this pack, if lingua-language-detector is missing the error prints the exact Python interpreter to install it into (the python_embeded problem strikes again). If you get "no videos returned", widen the window or the query; if you get "all fetched videos were filtered out", loosen min_views or exclude_keywords. And keep in mind this is an idea generator, not a fact engine - the scored "topics" are cleaned-up YouTube titles, so treat them as creative prompts for your story pipeline rather than guaranteed hits.