Use this live run sheet for the fixed two-hour session: 60 minutes of facilitated lesson, 30 minutes of student work, and 30 minutes for the guest.
Need the words to say and detailed contingencies? Open the evidence-informed Session 2 Live Presenter Script. It adds speakable transitions, safety notes, studio pathways, pacing flexes, and the guest handoff behind this run sheet.
Core question. When a model sees or generates an image, what is it actually working with?
Conceptual sequence. Pixels → human-created labels → shared text/image representations → diffusion.
Materials
- Session 2 live deck
- Session 1 recap
- The Squint Test (feature extraction)
- Image–Caption Match Lab
- Human Diffusion Canvas
- Diffusion Step-Through Viewer
- Default Test Comparison Viewer
- Prompt Pressure / CFG Scale
- Latent Space Compressor
- Image Prompt Pack
- Image Default Test Board
- AI Use + Consent Checklist
- CLIP project overview / Image–Caption Match Lab teaching move
Run of show
| Time | Segment | Facilitator move | Participant action |
|---|---|---|---|
| 0–5 | Welcome & bridge | Repeat consent and connect text choices to image choices. | Choose a participation pathway. |
| 5–10 | Session 1 retrieval | Share two artifacts or revisit the recap. | Name the investigation loop. |
| 10–17 | Pixels, labels & shared representations | Raise detail slowly, name PoseNet and the WordNet/ImageNet label layer, then use the Image–Caption Match Lab to predict, reveal, and revise one caption match. | Identify the first visual cue, justify one caption match, then change one phrase. |
| 17–25 | The room’s default | Tally first pictures for “a doctor,” then compare authored teaching simulations. | Treat each simulated scene as an illustrated hypothesis and name the documented real outputs needed for a model claim. |
| 25–36 | Diffusion — performed, then watched | Run the Human Diffusion Canvas, then the step-through viewer. | Draw an analogy by hand and predict what stabilizes next. |
| 36–44 | Guidance & revision | Compare vague, specific, and responsible revisions. | Distinguish prompt wording from CFG strength. |
| 44–52 | Whose picture? | Map user, system, and hidden human contributions. | Name one credit, consent, or disclosure decision. |
| 52–60 | Mechanism debrief & studio launch | Synthesize the loop and launch pathways. | Share one evidence-based claim. |
| 60–90 | Student work studio | Protect 25 minutes for investigation and five for posting. | Complete one observation and next test. |
| 90–120 | Guest spotlight, Q&A & close | Give the guest the full final block. | Track one visible and one hidden decision. |
Use the A/B/C comparison to compare a vague prompt, one added detail, and a responsible revision. Export only examples reviewed for consent and attribution. Canvas trajectories and debrief shares may be posted to the Evidence Wall only after affirmative sharing consent.
For credit, attribution, and recap decisions, use the existing AI Use + Consent Checklist rather than creating a new worksheet.
Facilitator prompts
- “What is one thing the Squint Test showed—and one thing it did not show?”
- “Which caption would probably sit closest to this image in a shared representation space—and what association makes it closer?”
- “When did composition commit on your canvas, and when did it stabilize in the viewer?”
- “What did the prompt leave unspecified, and what repeatedly filled the gap?”
- “What did guidance improve, and what did it break?”
- “Which human decision had the most power over the final image?”
- “What evidence would you need before making a claim about the model rather than one output?”
Investigation prompt
Choose a vague visual prompt such as “a doctor,” “a classroom,” or “a beautiful home.” What does the system fill in without being asked? Which details are technical defaults, and which are social defaults?
Low-AI / No-AI pathway
Participants can analyze curated screenshots or facilitator-provided image sets instead of generating images. The canvas analogy can also run with paper, thick-to-thin pens, and tracing-paper veils. Participants may design an age-appropriate Default Test without using AI tools.
Fallback plan
- If no image generator is available, use pre-generated examples or ask participants to predict likely defaults before revealing examples.
- If the Image–Caption Match Lab stalls, use the deck’s cat image and collect the A/B/C vote in chat; name any discussed ranking as an authored teaching example, not a CLIP result.
- If the Human Diffusion Canvas stalls, use paper and tracing-paper veils or collect each step’s proposed commitment in chat.
- If the Diffusion Viewer feels abstract, pause at seven named stages and ask participants to describe only what they can see.
- If Default Test prompts feel socially loaded, use safer prompts first, then explicitly frame more sensitive prompts as optional critique.
- If participants object to image generation, shift to consent, attribution, and visual-default analysis.
Live QA notes
During the session, note:
- Which image types in Feature Extraction were most useful.
- Whether the detail slider made recognition thresholds visible.
- Whether one predict → reveal → revise cycle made shared image/text representations clearer without being mistaken for image search or a real CLIP result.
- Whether the Diffusion Viewer explained denoising without live AI.
- Whether participants understood where the Human Diffusion Canvas analogy breaks.
- Which Default Test prompts produced the richest discussion.
- Any discomfort around bias, labor, consent, or artist imitation.