mediaai
Local AI media factory — Wikipedia → cinematic images, motion videos, spoken articles
—AI images
—audio clips
—motion videos
—on S3
Live gallery — latest generations
Loading live snapshot…
What it does
mediai is a headless pipeline that turns Wikipedia's most-viewed articles into a huge topic-varied library of cinematic media — one asset per subject, generated locally on an 8GB laptop GPU. Every image is anchored to a real Wikipedia photo via SDXL img2img, every audio clip reads the actual article summary, motion videos are tracked as an independent library fed by archive and new visual sources.
Built for the AI-VJ use case: an on-demand library of clips playable in the browser, with each clip tied to a real subject you can actually explain.
The pipeline
Wikipedia top articles -> prompts.jsonl -> article_worker
article_worker -> image_worker (SDXL image assets)
-> piper_worker (spoken-summary audio assets)
-> ltx_worker (optional AI video assets)
existing media archive + new visual sources
-> motion_worker (independent motion-library queue)
all completed assets -> S3 + MongoDB + GitHub index + Wikai BO
Nine workers, one supervisor
imageSDXL Juggernaut-XL img2img from Wikipedia photo, denoise 0.5, 30 steps, 1344×768 native 16:9
motion-imagesIndependent motion-library stream using archive and new visual sources; not gated by image-target progress
piperLocal neural TTS reading the actual Wikipedia article summary — ~1s/clip on CPU
articleFetches Wikipedia article text + hero photo per topic, backfills continuously
syncUploads every completed asset to s3://com27/media-baseline/ every 60s
mongoUpserts AIDB.media_baseline with searchable metadata
githubCommits per-asset JSON sidecars to flexappdev/media-baseline-index every 5 min
wikaiboRegisters each article in Wikai BO with wiki-rank-based priority score
youtubeNightly batch upload (500/day quota) — activates on OAuth token
The stack
Design principles
Local firstEvery step runs on your own hardware — no cloud API calls for generation. Zero token cost after setup.
Anchored to realityEvery asset ties back to a real Wikipedia article. No hallucinated "SDXL's imagined Paris" — actual Paris.
Restart-safeSQLite state + filesystem outputs. Kill any worker, restart supervisor, resumes exactly where it stopped.
Four storage destinationsS3, MongoDB, GitHub, YouTube — parallel sync workers, no single point of failure.
Rate-limit awareWikipedia backfill respects 429s; sync worker batches; workers heartbeat for supervisor auto-restart.
Person filterRejects >8000 living-person titles before generation to avoid deepfake territory.