Machines that move
When the model adds time, what starts to drift?
PoseNet can locate keypoints in one frame; CoTracker can follow selected points across frames. Generative video faces the harder inverse problem: creating each next frame while preserving coherence in identity, setting, and motion.
A video begins as still images ordered in time.
PoseNet turns one frame into estimated body keypoints; other models represent edges, texture, objects, or regions.
CoTracker follows selected points through a sequence, making “where did this feature go?” a concrete computational question.
A generator must create new frames while preserving identity, layout, motion, and physics; failures reveal where that relation breaks.
Teaching move · PoseNet → CoTracker: PoseNet can locate an elbow in one frame; CoTracker asks where that selected point went in later frames. Try that distinction in the Point Correspondence Lab, then name the harder generative problem: creating each new frame while keeping the elbow, the person, and the scene mutually coherent. Tracking analyzes correspondence—it is not itself video generation.
Build the same five-frame motion idea twice in Coherence Animator, changing only which visual references remain available.
Analyze pre-generated clips or still sequences for drift, physics breaks, and camera jumps.
Use Coherence Animator, paper, whiteboards, or still frames to compare previous-only with anchor-plus-previous drawing.
Prototype a frame viewer, overlay, or annotation tool for temporal failure modes.
Critique video as evidence, consent object, classroom artifact, or synthetic record.
Plan, not historical record. The July 25 live session followed a different path: assignment reviews, Session 2 feedback, the Coherence Animator and tool discussion, and Dr. Emily Thomforde’s talk. See the recap and recording chapters above for what happened live.
Welcome + two artifact shares
Re-enter through the mechanism map and three-line camp argument. Respond first through a private note or poll, then hear two pre-arranged shares—including one classroom-facing example: one artifact, one claim, one boundary each.
Active synthesis + point correspondence
Retrieve the Session 1 and 2 findings through private writing, a poll or precise chat prompt, and two voluntary responses. Then predict and reveal one clean track; distinguish tracking existing frames from generating new ones.
Coherence Animator · solo A/B comparison
Draw one shared opening and four new frames per run—nine drawings total. Run A shows only the previous frame; Run B shows the shared opening anchor plus the previous frame. Choose one feature, play A twice and B twice at the same speed, then name what changed and what the activity cannot prove about a real model.
Curated failure hunt
Track one feature at a time across two frozen examples. Ask four reusable questions: what changed, what supports it, what would verify the source, and what can we not conclude?
System map
Map model, interface, training data, workers, prompter, subject, editor, and platform. Keep consent and provenance separate, and turn the four questions into a reusable classroom protocol.
One-tool studio
Start from a completed worked example. The default frozen route ends with an observation, exact evidence, bounded claim, missing source evidence, and one “With learners, I would…” sentence.
Two claims + guest introduction
Hear two concise participant claims, then introduce Dr. Emily Thomforde and the listening question.
Guest talk + Q&A
Protect 18 minutes for the guest talk and eight minutes for participant questions.
Guest takeaway + close
Ask for one sentence to carry forward, return to fluent ≠ true / plausible ≠ neutral / smooth ≠ evidence, and invite the optional Studio.
Participants can complete the session with Coherence Animator, curated clips, or the same nine-drawing A/B sequence on paper. Temporal Telephone remains an optional facilitator-led group version. Direct video generation is optional. Open the full No-AI pathway.