Type one sentence — and ~10 minutes later you get a voiced, subtitled, 9:16 short film for a few dollars of compute. An autonomous showrunner on Qwen Cloud, with one human approval gate before any money is spent.

Inspiration

AI video tools render first and fix later — burning the most expensive step over and over. A real production locks the script and storyboard on paper first, then shoots once. We built that.

What it does

  • Trends → loglines: a live ribbon of trending topics, refreshed hourly, rewritten by the model into film ideas.
  • Writes the screenplay (qwen3.7-max), plans every shot (qwen3.7-plus).
  • Critic-before-render: a director-critic loop rewrites weak shots while everything is still text — ~$0.002 a round instead of ~$1 a re-render.
  • Paints the storyboard (qwen-image-2.0-pro); a series bible keeps characters consistent across shots.
  • Approval gate: redraw any frame with a note, reorder, drop — the only human step, before a single video credit is spent. The Film-it button shows the real price.
  • Films, voices, cuts: each approved still becomes a happyhorse-1.1-i2v clip, every line gets a voice (qwen3-tts-flash), ffmpeg cuts the film — with a caption and hashtags ready to post.
  • Cost ledger: every API call is metered; the price on the premiere screen is the real total, not an estimate.

How we built it

A zero-framework Python service on Alibaba Function Compute drives the whole pipeline over DashScope. One glass page: composer → live studio stream (script → board → critic → storyboard) → premiere.

Challenges

  • Image moderation false-positives ("hug") → G-rated symbolic staging; still blocked → shown honestly with a Regenerate button.
  • The video model can "succeed" with an all-black clip → an ffmpeg luma guard catches it, retries once, then refuses honestly.
  • Voice longer than its 5s clip → the cut holds the shot until the line finishes.

What's next

Shareable per-film pages, music/SFX bed, and a series mode with recurring characters.Type one sentence — and ~10 minutes later you get a voiced, subtitled, 9:16 short film for a few dollars of compute. An autonomous showrunner on Qwen Cloud, with one human approval gate before any money is spent.

Inspiration

AI video tools render first and fix later — burning the most expensive step over and over. A real production locks the script and storyboard on paper first, then shoots once. We built that.

What it does

  • Trends → loglines: a live ribbon of trending topics, refreshed hourly, rewritten by the model into film ideas.
  • Writes the screenplay (qwen3.7-max), plans every shot (qwen3.7-plus).
  • Critic-before-render: a director-critic loop rewrites weak shots while everything is still text — ~$0.002 a round instead of ~$1 a re-render.
  • Paints the storyboard (qwen-image-2.0-pro); a series bible keeps characters consistent across shots.
  • Approval gate: redraw any frame with a note, reorder, drop — the only human step, before a single video credit is spent. The Film-it button shows the real price.
  • Films, voices, cuts: each approved still becomes a happyhorse-1.1-i2v clip, every line gets a voice (qwen3-tts-flash), ffmpeg cuts the film — with a caption and hashtags ready to post.
  • Cost ledger: every API call is metered; the price on the premiere screen is the real total, not an estimate.

How we built it

A zero-framework Python service on Alibaba Function Compute drives the whole pipeline over DashScope. One glass page: composer → live studio stream (script → board → critic → storyboard) → premiere.

Challenges

  • Image moderation false-positives ("hug") → G-rated symbolic staging; still blocked → shown honestly with a Regenerate button.
  • The video model can "succeed" with an all-black clip → an ffmpeg luma guard catches it, retries once, then refuses honestly.
  • Voice longer than its 5s clip → the cut holds the shot until the line finishes.

What's next

Shareable per-film pages, music/SFX bed, and a series mode with recurring characters.

Built With

  • alibaba-cloud
  • dashscope
  • ffmpeg
  • function-compute
  • python
  • qwen
  • qwen-cloud
Share this project:

Updates

Submission history