The pitch is not a better model, it is fewer handoffs. An agent native to the canvas takes a plain-language brief, drops asset nodes onto the canvas, builds a nine-panel storyboard, and hands that grid to a video model — which is the same grid-to-video bridge that shows up in the author's LTX work, only automated.

Evidence note: this episode ships no workflow JSON and 说明.txt records no workflow or mirror link. This is a walkthrough of a closed platform surface; the steps are the author's, the numbers are his spoken settings.

When this is useful

  • Going from a one-line idea to a first cut without learning a node graph.
  • Producing a storyboard grid that constrains a video model's timing before you spend on generation.
  • Building a short piece from several 15-second segments stitched by continuation.
  • Upscaling the low-resolution video output economically.

How this differs from the previous build

The change is where the orchestration lives.

  • The agent writes the graph, not you. Ask for a nine-panel storyboard and the workflow materialises on the canvas.
  • The storyboard is the contract. The author is explicit that skipping it makes the video step drift, even though the model is strong enough to go straight from assets to video.
  • Continuation replaces re-generation. Longer pieces come from feeding the previous 15-second clip plus assets back in, not from one long prompt.
  • Upscaling is a separate, cheap step because the video model outputs at 480p.

Running it

  1. Generate assets first. The demo asks for a Chinese-style mecha in three-view form, then a dragon.
  2. Pick the image model per asset and set the aspect ratio — 16:9 in the demo.
  3. Refine the prompt when an asset comes out malformed, and regenerate.
  4. Upload the assets and ask the agent for a nine-panel storyboard. It creates the workflow on the canvas.
  5. Generate the grid at the highest resolution you can justify; the author uses a 4K 16:9 pass.
  6. Drag a connector from the grid to the video node. Feed the grid plus a detailed prompt to the video model.
  7. Write the prompt with explicit per-second beats — 0-1s, 1-2s, through to the last second — and include style and audio direction.
  8. For longer pieces, continue from the previous clip and add assets; then upscale the 480p result.

Honest limits

  • The video step is the most expensive part of the flow, and the model outputs at 480p before any upscale.
  • The author still recommends building the storyboard first; skipping it lets the video drift even with a strong model.
  • Anything the agent generates is platform-managed — there is no portable graph to take elsewhere.
  • No workflow JSON and no link are recorded in 说明.txt for this episode.
  • The per-second prompt discipline is manual work; the agent does not write it for you.

Episode resources