Web Media Getter

web-media

Query many free image/video sources in one fan-out, get a normalized result list, optionally download top-K with an attribution sidecar. Zero-dep stdlib script.

Script: webmedia.py (in this dir). Keys: PEXELS_API_KEY, PIXABAY_API_KEY in central/.env (optional — the 5 no-key sources work without them).

Sources

GIF sources fire only with --type gif. Keys: KLIPY_API_KEY, GIPHY_API_KEY in central/.env. (tenor adapter removed — Google EOL'd the API 2026-06-30.) klipy is the one to get (free + unlimited); its adapter is unverified — assumes Tenor-compatible request/response; verify against docs.klipy.com when you key it. webmedia.py "shrug" --type gif --count 6 --json

Usage

webmedia.py "1950s street scene" --type image --count 8 --json
webmedia.py "rocket launch" --type video --source nasa,internetarchive
webmedia.py "car factory 1930s" --source all --download --out /tmp/cars

• --source all (default) | nokey (no-key only) | comma list (wikimedia,pexels)
• --type image|video · --count N · --json · --download --out DIR
• --download fetches each result's direct media URL and writes attribution.json (source, author, license, url, page_url) alongside the files.

Record schema

{source, title, url, thumb, dl, page_url, author, license, w, h, type} — dl is the directly-downloadable media URL (None when only a page exists).

The video caveat (important)

Archival sources (Internet Archive, Europeana, LoC) host whole films/documentaries, not single shots. So:

• Modern single clip → pexels / pixabay (born as short clips, direct MP4). Done.
• Historical single shot → retrieve the IA film here, then extract the shot:
  • Twelve Labs Marengo search (free 600 min) — pass the IA public MP4 URL, get a timestamped moment for "car on assembly line", clip with ffmpeg. Semantic, cheap.
  • or PySceneDetect (free, local) to cut the film into shots, then rank keyframes with CLIP via the muser skill. Fully offline.

Audio: freesound + audio QA

webmedia.py is image/video. For sound effects (real, CC-licensed) and for judging audio (since Claude can't hear), two sibling scripts live in central/scripts/:

• freesound-fetch.py "<query>" [count] [max_sec] [out_dir] — searches freesound.org and downloads short hq-mp3 previews. Prints one JSON line per file with license/user for attribution. Key: FREESOUND_API_KEY in central/.env (token-based read; full originals would need OAuth — previews suffice for SFX).
• audio-judge.py <file> "<target>" — sends the clip to OpenAI gpt-audio (audio-native) and returns JSON {heard, score, matches, suggestion}, enabling a generate/fetch → judge → iterate loop. Auto-sources a real sk- OPENAI_API_KEY from .env (ignores a local lm-studio stub env var). Pads sub-2s clips so the speech-tuned model doesn't refuse. Caveat: it reliably describes audio and filters obvious mismatches, but it is NOT a trustworthy judge of subjective qualities like "grating" — it labels nearly any beep "sharp/high-pitched". Use it to cull, not to make the final aesthetic call; confirm by ear.

Where this fits

This is the internet-retrieval capability — peer to muser (local semantic search) and fal (generate). A future media router would fan out across all three and rank candidates by relevance (CLIP), handing aesthetic spreads to lookdev. Don't build that router until the model demonstrably mis-routes without it.