# Oct 16, 13:30

## Overview

Over lunch at Suzy's, two filmmakers walk through how one of them made a quick AI test commercial for the roster in about an afternoon, start to finish. They cover the steps: script, music, voices, sound effects, character faces, storyboards and animatic, LTX renders, editing in DaVinci Resolve, and a final upscale. The commercial features Dave, the boss and a roster creative. The other speaker compares it with their own workflow and probes the order of steps, cutting, and upscaling. What came out of it: the commercial will become a case study (and suits a Sherpa), the speaker will share their lyric-removal skill, and they agreed to test whether 4K references beat upscaling the finished video.

## Key points

### Background and speed
- The first commercial, for Ring My Case, took a couple of weeks while the skills were still being built.
- The next one used the client's real footage and voices, took 30 billed hours, and earned $6,000.
- The newest test began after lunch, partly driven from a phone: a URL went into Story Circle, and the story analyzer scored it in the 90s, then pushed it toward 100.
- A skill now analyzes a QuickTime file both by audio transcript and visually, producing a contact sheet of key frames. A transcript alone would miss nuance.

### Story first, then art
- The StoryBrand framework gives the "hooks" (excite, disturb, assure). Pick the hero; the brand is the guide ("the Yoda").
- In the commercial, the boss is the hero, Dave is the comic relief who says "we need professionals," and the roster creative is the guide. The roster's aim is to make the customer the hero.
- Story structure gives the beats; emotion, camera, timing and look are "the art," added afterwards.
- The first pass came back dry. The fix was to feed in the latest Apple commercial (the "baby iMac" ad) and a cinematic/filmic prompt.
- Color subtext: the break room starts cold and bluish, and the scene ends warm once the roster creative (in warm colors) arrives.
- A blind test with Veronica: she rated the plain version good, but saw the difference in emotion, line delivery and color in the finished one.

### Voices
- Daisy bird's 100 voices sounded robotic and flat on their own.
- Voiceify reworks lines to sound human. It grew from Chipify, which learned the speaker's cadence from their YouTube videos.
- Voiceify works one sentence or paragraph at a time, with surrounding context, and takes emotion directions (questioning, scared, screaming).

### Order of work
- The order: script, then music, then voices, then sound effects, then character faces, then storyboards and animatic, then video renders.
- Music comes before voice: "the music is all the emotion," and editing cuts to it. Music comes from Suno and YuE, plus a skill that strips lyrics.
- The same lyric-removal skill pulled the voices out of the Apple ad to study its music and the "click."
- Sound effects come from Sound Snap (a Hollywood library). A new open sound API (55.2 million files) from today's dailies now serves as a backup source. The AI places, levels and normalizes the effects in the timeline.
- Prompts take on roles (Hollywood Foley artist, cinematographer, voice) to read the script.
- No prop or location sheets were made, only character heads from Nana Banana Pro, first take. The characters held steady through prompting and start frames.
- Storyboards act as Omni references for the video renders, not exact first frames.

### Rendering and editing
- Rendering ran locally on LTX, about 15 minutes for under six seconds. Prompts were written for LTX: short, efficient, using camera moves taken from the Apple ad, such as a slow push on Dave and a gimbal wraparound from the boss to Dave.
- The other speaker's renders are "all on sticks," with no camera movement, which loses the power of film language.
- Everything goes into DaVinci Resolve from the animatic on. Claude handles track fixes, ducking (for example, raised from 0.5 to 0.875 per section) and redoing single lines ("clip 17, sound more scared").
- Editing rules: cut to rhythm, use J-cuts, "start late, leave early," and don't hold shots too long (a note for Eric).
- The ending moved to a second, different music track. The music drops out at the boss's low moment.
- Fast LTX renders could replace the client sign-off on animatics, which once guarded against a "$4,000 shot."

### Upscaling
- The finished cut is a "comp." For the real version, go back to the key frames, upscale them to 4K as Omni references, and re-render without changing the story.
- Feeding a 4K image holds detail, such as logos on shirts, even after the model downsamples it. Eric works in 4K images, renders at 1080 and upscales to 4K.
- Topaz is used locally for images. The LTX output was about 944, not 1080.

## Action items
- Build a case study of the commercial, with the saved beats and files.
- Send the lyric-removal skill to the other speaker.
- Show the character sheets/heads and how they were made, at the house.
- Show the upscale workflow when back from lunch.
- Test 4K Omni-reference re-renders against upscaling the whole video.
- Check the LTX render resolution.
- Look into stringing short H3 clips together.
- Put together an image upscaling pipeline.
- Review the open sound-effects API sent to the team.
- Rewatch the commercial with attention to the music.
