The pipeline is a real film production, compressed
A film set doesn't point a camera at actors and hope. It breaks down the script, casts, scouts locations, storyboards, blocks the scene, then rolls. Shows follows the same order — and for the same reason: every phase catches mistakes that would be 10x more expensive to fix in the next one.
Script → Breakdown → Canon → Storyboard → Staging → Generate → Export
free free 1cr/img free 1cr/img 1-4cr/sec free
Phase by phase
Script (free)
Your input surface. A screenplay, a story chapter, a rough scene — anything with characters, places, and events. You also choose the format (short film, ad, UGC, trailer, narrated story), which changes how the director-brain plans coverage, and the model cast (more in lesson 6).
Breakdown (free)
The LLM reads your script like a 1st assistant director: cast list, locations, props, a director's plan, and per-scene shot lists with dialogue assigned line by line. Long speeches get split across multiple shots automatically — stuffing a paragraph into one AI video clip wrecks quality, so the system never does it.
This is also where you get the budget estimate: canon images + storyboard frames + every shot priced at its model's rate + a suggested 20% retake buffer. Nothing has been charged yet. If the film is too expensive, cut shots or switch models here.
Canon (1 credit per image)
Characters, locations, and key props each get one locked reference image. Everything downstream generates from these images. This is the consistency engine — skip the approval step carelessly and you'll pay for it in every later shot.
Storyboard (free)
The plan, laid out shot by shot: framing (WIDE / OTS / MCU / CU), the prompt the system wrote, the dialogue covered, duration, references, and the rendering model. Edit freely — this is the last all-text, all-free phase.
Staging (1 credit per frame)
Every shot is composed as a still before any footage. Angles chain off each other so the space holds (lesson 4 is all about this). This is the cheapest place in the whole pipeline to burn retakes — spend your budget here, not on video regenerations.
Generate (1–4 credits per second, by model)
Each approved still gets animated. Every render is a take; retakes stack up and you can hot-swap between them forever. Approve / Edit / Regenerate / Iterate until picture lock.
Export (free)
Numbered clips for your editor, or a server-assembled preview cut with voiceover and music mixed in, plus a 9:16 social render and .srt captions.
Where credits actually go — a worked example
The Last Transmission (≈12 shots, ~70 seconds, LTX Fast):
| Item | Cost |
|---|---|
| 4 canon images | 4 cr |
| 12 storyboard frames | 12 cr |
| ~70s of video @ 1 cr/s | ~70 cr |
| Planned total | ~86 cr |
| Suggested retake buffer (20%) | ~17 cr |
Notice the shape: video is ~80% of the cost. Every mistake you catch at the still stage (1 cr) instead of the video stage (5–10 cr per retake) is money back in your buffer.
The three rules of not wasting credits
- Never generate video from a still you haven't looked at. Staging exists so you don't have to.
- Fix identity problems at Canon, not at Generate. If a character drifts, the fix is almost never "regenerate the shot" — it's "fix the canon image the shot inherits from."
- Use Iterate before the third Regenerate. If two takes in a row miss the same way, the prompt is the problem. Tell the director what's wrong and let it rewrite.
Next lesson: Characters that stay consistent — canon locking and wardrobe variants in depth.