test

Overview

The speaker talks with Grock about the design of Scribeboard, a tool that turns a story synopsis into a hand-drawn whiteboard-style video. They fix a flaw in the plan: each scene’s image must match the scene’s length. Drawing speed stays the same everywhere, and longer scenes get richer, more detailed pictures. So images now come after the voice track. They also add a quick virtual unmasking test, choose styles in modules (RSA Animate and SketchNote first), and pick the resolution once at the start. Last, they fix the resolution model: generate each scene image at ChatGPT’s highest resolution, upscale it with AI to 4k, stitch the images onto one large canvas, and have FFMPEG render a final 1920 by 1080 video. Grock checks the name, and “Scribeboard” and Scribeboard.com both look free. Grock sends each change to Wipe, who revises the outline.

Key points

Pacing and image complexity

  • Drawing speed stays constant across every scene; a longer scene gets a more complex image, not a slower drawing.
  • A complexity budget estimates strokes per second and tells the image generator how dense each scene’s art should be.
  • Images now generate after the voice track is recorded and aligned, because density depends on the real narration length.
  • The speaker supplies only a story or synopsis; the AI writes the narration and decides the scene count, length and content.
  • Images that come out too dense or too sparse get regenerated.
  • New Step 7: a virtual unmasking test replays the strokes at standard speed against the real time code. It checks that drawing finishes on time with the right hold, that nothing stays undrawn, and that the camera hits the right spots. If the test fails, the image is regenerated or the stroke order fixed before the full render.

Styles and resolution

  • Styles become modules the user can choose from; RSA Animate and SketchNote come first.
  • A project setup step picks the style and the resolution up front. Script vocabulary, image density, segmentation rules and animation all read from the chosen style module.
  • One resolution choice sets the canvas size, the output video size and the per-scene image size.
  • Corrected model: the final video is 1920 by 1080. Each scene image is generated at ChatGPT’s highest resolution, upscaled with AI to 4k, then placed on one large working canvas.
  • Masking and stroke order stay deterministic, driven by the segmentation data. AI may control the masking, but it generates no new pixels.
  • FFMPEG slides the viewport across the canvas and reveals strokes in order. It handles 1080p output, but could choke on a 20k or 30k image.
  • Grock noted that current image models top out around 1500 by 1000 pixels, which led to the upscaling plan.

Product name

  • “Scribeboard” as one word appears to be free; Scribeboard.com resolves to nothing.
  • Close names: Logitech’s Scribe whiteboard camera, Noel’s Scribe mobile marker board, a 2005 Mac app called Scribeboard, and Scribe Animator (the closest in spirit).

Action items

  • Register Scribeboard.com before someone else does.
  • Find an AI upscaler that doesn’t distort the images too much.
  • Review Wipe’s updated outline with the corrected resolution model.