# test

## Scene Image Pacing Flaw

### Images Must Match Scene Pace

- Original scene images must be generated to the pace of each scene
- Scene pace set by the recorded voice
- Image generation must judge complexity from pace
    - Example: six scenes; one 20 seconds, one 40 seconds
- Same drawing speed for every scene; longer scene gets more complex image
    - Don't slow the rendering down

### Grok's Response

- Current spec assumes fixed drawing speed, stretches or compresses to fit
    - Short scene feels rushed, long one sluggish
- Image generator should take scene duration as input
- Scale complexity to duration
    - More elements, strokes, detail for 40-second scene; fewer for 20-second
- Every scene draws at same natural pace
- Proposed complexity budget in script phase
    - Estimates stroke count per second
    - Tells image generator how dense each scene's art should be

## AI Control of Narration and Scenes

- User writes story or synopsis; AI writes final narration
- AI fully controls scene breakup and scene content
- Synopsis only the seed
    - AI decides scene count, scene length, what gets drawn, image complexity
- Full creative control over pacing and density, for hands-off pipeline

## Spec Update Request

- Insert into spec; skip technical spec, keep basic algorithm only
- Update sent to Wipe; outline revised with complexity scaling rule

## Product Name

- Name confirmed: Scribeboard

## Revised Outline Flow

- Big change: images generated after voice track, not before
    - Image density depends on real narration length, known only once audio exists
- Flow:
    - Synopsis in
    - AI writes script, decides scene count and links
    - Voice recorded and aligned
    - Complexity budget sets image density from actual drawing time
    - Images generated to budget; too dense or too sparse ones regenerated
- Drawing speed constant across scenes; longer scenes get richer pictures

## Virtual Unmasking Test

- Proposed: quick virtual unmasking test after each image, against time code
- Fast dry run, no full render
    - Checks drawing finishes on time
    - Checks nothing left undrawn
    - Checks camera hits right spots
- Failures caught cheaply, regenerated before expensive full render
- Added to outline as Step 7
    - Simulated playback reveals strokes at standard speed against real time code
    - Checks drawing finishes with right hold
    - On failure, loops back to regenerate image or fix stroke order

## Scribeboard Name Check

- Goal: avoid taking an existing name
- "Scribeboard" as one word appears untaken
- Close cousins:
    - Logitech Scribe: whiteboard camera for video conferencing
    - Noel Scribe: mobile marker board
    - Scribeboard: old Mac app from 2005 for drawing on screen
    - Scribe Animator: whiteboard explainer animation studio; closest in spirit
- Scribeboard.com resolves to nothing
    - No active site, no parked page
    - Looks available; advice to grab it

## Modular Styles

- Styles should be modular
- RSA Animate style already discussed
- Sketch note style as another option
    - Different from RSA Animate
- Aim: several styles to choose from

## Final Resolution Choice

### Resolution as One Decision

- Project needs a choice of final image resolution
- Style: sets drawing speed and reveal moves; picked at project start
- Resolution one decision, not three
    - Example: pick 4K output
    - Sets canvas size, video resolution, per-scene image resolution at once
    - Every image dense enough for final canvas

### Updated Outline from Wipe

- Project setup step: pick style and resolution up front
- Style module library; RSA Animate and SketchNote first two
- Resolution plan computes canvas, output video, per-scene image size from one choice
- Downstream steps read from chosen style module
    - Script vocabulary, image density, segmentation rules, animation behavior

### Video Resolution Drives Image Size

- User confirms: choice is final video resolution
- Final image used much larger, but calculated from video resolution
- Grok agrees
    - Canvas three to four times wider than camera close-up view
    - Source images on canvas much larger than 4K
    - Every pipeline size traces back to that one number

## Image Model Resolution Limits

- Question: will Claude or Codex have a problem?
- Grok: yes, real bottleneck
    - Current image models top out around 1500 by 1000 pixels
    - 8K canvas, scene covering a quarter: roughly 4000 pixels wide
    - No model does that natively today

## User's Correction: Upscale and Composite

- Six scenes means six images
- Render each at highest ChatGPT resolution
- AI-upscale each to 4K
    - Pick an upscaler that doesn't distort much
- Stitch upscaled images together
- Masking deterministic; AI controls it but generates no image
- Video masking done in FFMPEG
- Destination should be 1920 by 1080, final video size
- Six-scene canvas still very large
    - FFMPEG can handle that, but not a 20K or 30K image

## Grok's Restated Design

- Each scene image at max ChatGPT resolution, upscaled cleanly to 4K
- Composited onto one big working canvas
    - Could be 16K or whatever layout needs
- Masking and stroke order fully deterministic
    - Driven by segmentation data, not new pixels
- FFMPEG does video work
    - Slides viewport across canvas
    - Reveals strokes in order
    - Outputs final 1920 by 1080 video
- FFMPEG handles 1080p easily
- Big canvas is source material, not output
- Correction sent to Wipe; resolution model flipped, updated outline to follow
