Once you can make an image, you can edit it, move it, and lift it off the page.
By the end of Day 2, students understand that AI literacy is partly about control: knowing what to change, what to preserve, and what a model is likely to invent. They use masks, outpainting, motion prompts, and 3D conversion to test those ideas. The artifact is a record of the experiment, but the real learning is being able to explain why a small constraint changed the result.
Schedule
Block 1: instructor demo, class co-construction, student studio time
Block 2: second lesson sequence or guided production sprint
Break: 15 minutes away from the screen
Block 3: studio work with instructor/tutor nudges
Block 4: finish, export, document
Daily share-back: every student shows one decision
Reflection card and next-day preview
Lessons
Lesson 2.1 - Edits and masks
Use surgical control: change a part, keep the rest.
I do / We do / You do
I do: 15 min: show inpaint, outpaint, and style-transfer from one source image.
We do: 10 min: class masks a region of one shared image and rewrites only that area.
You do: 25 min: each student edits one Day 1 image three ways.
Deliverable: Edit triptych plus one sentence, "what changed."
Stretch: Produce a four-step edit progression that tells a tiny story.
AI tutor aiGuidance excerpt
- nudge: Ask the student to name the region before changing it.
- nudge: If an edit changes too much, suggest a smaller mask or a more literal instruction.
- require: Do not upload or modify identifiable photos of classmates or minors.
Lesson 2.2 - Time or dimension
Lift a still into motion or volume.
I do / We do / You do
I do: 15 min: show one Wan 2.2 video path and one HunYuan3D image-to-mesh path.
We do: 10 min: run both paths from one class image and compare affordances.
You do: 25 min: each student chooses video or 3D and produces one finished artifact.
Deliverable: One looping 4-8s MP4 or one rotatable GLB. If HPC is down, video is instructor-demo only and students default to 3D or local queue.
Stretch: Video students try a second clip with a different motion prompt; 3D students try a texture override.
AI tutor aiGuidance excerpt
- nudge: Ask whether the artifact needs motion or volume to make the idea clearer.
- nudge: For video, ask for one camera move and one subject action, not a paragraph.
- require: For 3D scans or likenesses, use only objects or self-created imagery with consent.
How the tech works
Editing models make the iteration loop more local. Instead of asking for a whole new scene, the student names a region and describes what should change there. Functionally, the model keeps the surrounding context while replacing or extending a selected part. Video generation adds another constraint: frames have to stay coherent over time, so a useful prompt specifies one motion, one camera behavior, and one duration. Image-to-3D is different again: the system infers silhouette, surface, and texture from a still image, then exports a mesh-like object students can inspect. ij8 currently supports image, video, and 3D generation, with video best on HPC and HunYuan3D workable through the 3D pathway. The lesson copy therefore does not promise per-lesson model pinning or public preview links; it describes production choices inside the available studio. The practical difference for students is that edit work is about decomposition: separate the scene into parts, choose the part that needs intervention, and preserve what already works. The practical difference for the instructor is queue management. Video can be slow when HPC is unavailable, so the lesson keeps 3D as a parallel path and makes the fallback explicit instead of hiding it. Students still complete one artifact, and no one is sent into a different lesson track (Modern Classrooms self-pacing; NextGen self-pacing).
A short history
Digital artists edited pixels long before AI inpainting. What changes in the 2022 onward period is that the replacement can be generated from language while preserving local context. Video follows a similar public arc: experimental AI video systems became progressively more usable, and by 2024 tools like Sora and Runway Gen-3 made generated motion a general media story. 3D generation is newer as a student-facing workflow, but it connects to older computer-graphics problems: how to infer form, surface, and texture from limited views. The day's history is therefore a bridge between familiar image editing and more spatial forms of media. It also gives students a reason to compare media honestly. A still image can carry detail, a video can carry timing, and a 3D object can carry inspection and scale. The assignment asks them to choose the medium that serves the idea rather than treating every new output type as automatically better.
Ethics for this day
Day 2 raises the stakes because editing and motion can make fabricated media feel more evidentiary. Thorn's data and CNN's reporting on explicit deepfake abuse show why consent rules have to cover moving images, not just stills (Thorn research; CNN reporting). The 3D path adds another consent issue: scanning or reconstructing real people can create reusable likeness assets. Students work from their own generated images or objects, not classmates' faces. Provenance is introduced as a practical habit: keep the source image, prompt, edit notes, and final export together. That habit protects the artist as well as the audience. If someone asks how a clip or mesh was made, the student can show the chain of decisions instead of relying on memory. It also makes attribution easier on Day 4, when selected work moves from private studio output to public-facing portfolio material.
Real-world applications
Film pre-visualization and animatics
Game asset prototyping
Fashion lookbook motion studies
Architectural reference and massing studies
Museum installation mockups
Famous works to study
Refik Anadol, Echoes of the Earth: Living Archive (2024). Immersive AI installation at Serpentine using environmental datasets. Source.
Memo Akten, Learning to See (2017-18). Camera-based neural interpretation work that makes model perception visible. Source.
Runway, AI Film Festival (2023-). Festival platform for studying current AI moving-image practices without inventing Sora-era canons. Source.
Refik Anadol, Large Nature Model / Coral studies (2024-). Nature-data-driven AI research and installation work around environmental archives. Source.
What you should be able to do by the end of today
Edit one part of an image without destroying the whole scene.
Choose whether an idea needs time or dimension.
Export one MP4 or GLB through the available studio path.
Document source, prompt, and final artifact together.
Deeper look
Optional note for the curious
Video generators need temporal coherence: a chair should remain the same chair from frame to frame unless the prompt says otherwise. That is why short clips with one clear motion often work better than ambitious prompts. In 3D, the model is guessing unseen sides, so inspection and revision matter.