# Oct 16, 13:30

## Speed of the New Workflow

### Earlier Commercials
- Ring My Case commercial: skill-building, took a couple weeks to set up
- Last commercial: $6,000 from client, 30 hours billed
    - Used real recorded footage, real voices, real interview footage
    - Had to cut through that footage

### Yesterday's Commercial
- Started after lunch, driven from the phone
- URL into story circle, story analyzer run, score requested

## Story Analyzer Scoring
- Asked how to get storyline to 100
- Came back with higher score, 90-something
- Question: scored as a document, or as a video script?

## Video Analysis Skill
- Takes the QuickTime; audio transcript plus visual analyzer
- Produces a contact sheet
- Frames at key points, not fixed intervals
    - Significant action, camera lens changes
- Transcript alone would miss nuance

## Video as a Visual, Emotional Medium
- Video by nature visual; strongest component is emotion
- Told people at IBM: video carries an entertainment component
    - Corporate pushback: "this ain't Netflix"
- Talking head bores viewers if held too long

## From Concept to Video Script
- Usually start from script, treatment or story, not video
- StoryBrand framework gives hooks to hang story on
    - Excite, disturb, assure: understand the why behind the structure

### Designating the Hero
- No hero, no star; everyone roots for the hero
- Suzy example: customers are the hero, Suzy is the Yoda
    - Suzy gives the hero a good lunch

### Roles in This Commercial
- Boss is the hero
- Dave: not the hero, comedy relief
    - Screwed up; says "we need to get professionals"
- Hired roster creative: the guide, the Yoda
- Roster's intent: make the customer the hero

## Limits of Story Structure
- Story lacks video sense: emotion, camera, gravitas, timing, beats
- Structure nails hero, guide, plan and path
- Visuals are where the art comes in

## Art Direction

### Making It Human and Cinematic
- First pass gave a dry delivery
- Asked for more human; that is the art, not story direction
- Story gave all the beats; visuals needed to be cinematic

### Apple Commercial Reference
- Fed it the latest Apple commercial for lighting and look
    - Baby iMac, resembling a 90s iMac
- Wanted a particular look and tone

### Cold-to-Warm Subtext
- Break room cold, bluish
- Ending warm
- Roster creative arrives wearing warm colors

## Case Study Plans
- Process was spoken, not typed; beats saved
- Good case study, good one for a Sherpa
- Blind test with Veronica
    - Other version judged good; theirs showed emotion, line delivery, color

## Directing Look and Delivery
- Visual look, camera angle, camera movement
- Two inputs: Apple commercial plus own cinematic filmic prompt
- Audio: asked for the way humans talk
    - Arms, eyes; Dave's mannerisms

## Voices

### First Attempt
- Daisy bird API: about 100 voices
- Sounded good but robotic, flat delivery

### Voiceify
- Modified version of Chippify
- Run repeatedly: more human, more nuance
- Any voice can go through Voiceify

### Chippify Foundation
- Fed own YouTube videos from over the years
    - Cadence, spacing, timing, language
- Rewrites any text to sound like them, with timing
- Voiceify built by asking what makes voices human

### Line-by-Line Processing
- Breaks transcribed lines into pieces
- Filmmaking parallel: shoot in short takes, not 10-minute runs
    - Camera moves, lines delivered again
- One sentence or paragraph at a time, with surrounding context
- Emotion specified per line
    - Questioning, scared, screaming

## Characters (First Mention)
- One pass; first pass looked good
- Done in the wanted filmic quality

## Assets
- No props or location sheets this time
- Aimed for "the Auto-Tune of commercials"
- Components: script, characters, voices, Apple look and feel, lockdown
    - Apple-based look with cold and warm

## Music and Audio Assembly

### Voice Tracks
- Other speaker: renders one full voice track, then shots
- This workflow: individual audio files per line
- Everything goes into DaVinci Resolve
    - Music, individual audio tracks, videos all visible
- Tweaks become the new baseline

### Cutting to Music
- Uses J cuts: next scene's audio starts early
- Cuts to the music; music carries the emotion

### Sound Effects
- Soundsnap connected; Hollywood-quality library
- Dailies surfaced an API sound-effects library; added as secondary
    - 55.2 million audio files: effects, ambient, music
    - Same one the other speaker sent the team
- Searches, drops files into folder and timeline
- Sets levels, normalizes audio at engineer level

## Production Sequence
- Script first, then music, then voice
- Listens to music while writing the script
    - Spotify playlists
    - Tarantino influence; can't afford licensed tracks
    - Suno gets the feel
- Other speaker barely notices music; this speaker notices it first

## Sound Effects from the Script
- AI picks up setting: office, New York, urban
- Prompted to read script as a Hollywood Foley artist
- Same approach for cinematography and voice

## Music Tools
- Suno and YuE
- YuE mostly produces lyrics; a skill removes them
    - Skill to be shared with the other speaker
- Stripped Apple commercial's voices to study its music
    - Snap, click cues; part of the case study

## Rhythm in Editing
- Editing prompts know about rhythm
- Rhythm also carries subtext
- Horror: beat speeds as monster nears; Jaws
- Applies to commercials and corporate work too

## Characters

### Character Images
- Nano Banana Pro; first take kept
    - Dave, boss, CEO, creator all looked right
- Heads only, no full character sheets
- Meant as a quick-run test

### Holding Characters
- Some drift, yes and no
- Nano Banana Pro held characters via start frames
- Prompt told it to hold the character

### Casting Looks
- Other speaker: avoid too-beautiful people, an AI giveaway
- Commercial branding: real life with boring parts removed
- Typecasting; displacing a character makes comedy
    - Two gray-haired old guys at a bar; nursing-home pickup idea

## Storyboards
- Built after characters; script already locked
- Goal: animatic with music, voiceovers, faces
- Storyboards hit all the beats

### Storyboard vs. First Frame
- Other speaker: storyboard represents a shot, not a start frame
- Speaker: roughly the same
    - Like an on-set photographer's high-fidelity still
    - Camera may not start there
- Prompted only to create storyboards from script
- Used as Omni references, not first frames

## Rendering with LTX
- Rendered on own machine
- Other speaker: H3 supposedly better
    - Seedance comparison; rewrote using Eric's comments
- Render time: about 15 minutes for under six seconds
    - Other speaker: about seven minutes
- Weighted, compressed faster model

## Holding Detail

### 4K Reference Images
- Eric's method: render image at 4K, feed back in
    - Holds details like shirt logos
- Nano Banana image upscaled to 4K before video
- Downsampled 4K input still gives a better image

### Upscaling Tools
- Topaz, local, for images
- Other speaker: has both image and video; pipeline for video
- Upscales whole video, not 4K images
    - 1080p upscales can be too sharp
    - 24 fps for cinematic look

### Faces
- Ways to re-render faces in fast or distant shots
    - Motion blur, faces drifting at distance

## Shot Prompting
- Fed cinematic language, cinematic Bible, Apple commercial
- Prompted specifically for LTX; new, no prior prompts
- Prompt-for-action rules; character limits; fewer is better

### Camera Moves from Apple Analysis
- Slow push on Dave
- Wraparound starting on boss, ending on Dave
    - Follow the action, smooth gimbal move
- Other speaker's work "all on sticks": cut, cut, static
- Roster creative gets full-screen close-up
    - Yoda gets full-screen importance

## DaVinci and Animatics
- Renders drop into DaVinci automatically
- In DaVinci from music and audio onward
- Animatic: storyboards with music, sound, sometimes effects
- Pre-AI: animator drawings, client sign-off before hiring crew
    - Signed animatic settles disputes; "$4,000 shot"
- LTX speed may make animatics unnecessary

## Editing in DaVinci
- Always tweaks cuts once audio is in
- Music on separate tracks from voice
- Claude fixes audio after moves
    - Track A17; everything properly labeled
- Tweaks sound-effect levels and placement

### Music Switch and Ducking
- Switched final song late; two songs
- Ducking under voiceover done automatically
- First pass too stark; asked for softer, graduated
    - Values per section: 0.5 to 0.875
- Listens on headphones, then without

### Fixing Lines
- Redo clip 17 audio, "more scared"
- Fixed and placed back in timeline
- Other speaker prefers this to their own method

## Cuts and Pacing
- J-cut, L-cut; hearing next scene first
- Every frame counts; push the action
- Eric holds cuts too long
    - 1.5 seconds; cut a second off
    - Pretty shots that don't lead the story waste time
- Start late, leave early
    - Boss scene: audience only needs to know he's frustrated
- J cut under previous shot; audience fills it in

## Clip Length and Sync
- Other speaker's scripts time actions by the second
- LTX renders whole shot; trimmed in DaVinci
- Re-renders whole shot; cut timing unchanged
- Dave and boss clips trimmed early; full clips would overlap
- Voices on separate tracks; music independent, drives everything

### Music at the End
- Cut to a completely different track, similar feel
- Two music pieces instead of one long one
- Music stops when he's down on the phone
    - The "crap moment"; deliberate choice

## Comp and Final Upscale
- Rendered at 720p; result is a comp, not the real thing
- Next: back to key frames, upscale everything
    - Story unchanged; only visual quality raised
- Upscaling whole video loses quality; need 4K Omni reference
- Latent upscaling raised; agreed to test both

### Eric's Pipeline
- All characters, locations, props at 4K
- Renders 1080, upscales to 4K at the end
- LTX can't render 1080; H3 maybe, only two or three seconds
- LTX render was about 944
