LECTURE 02 · SEPTEMBER 8, 2026

Generative editing:
images, art, and video

90-minute lecture · Slides · English / 中文 transcripts

All discussion questions & suggested answers ↓

WATCH THE LECTURE · ENGLISH / 中文

One artwork. Three ways to think.

Explore the lecture with English and Chinese captions. Use the chapters below to revisit a topic.

SLIDE 001 OF 123Generative editing

Opening

1 / 123← → to navigate
Slide 1: Generative editing
LECTURE SECTIONSJump to a section
SLIDE 001

Generative editing

14 DISCUSSIONS · 42 QUESTIONS

Discuss, then compare.

Explore each question before opening its suggested answer. These are starting points for discussion; other interpretations are welcome when supported by visual evidence and clear reasoning.

Slide 8 →

Before Part 1: image editing

  1. When glass becomes ceramic, what should stay the same?

    Suggested answer

    Preserve the pear’s silhouette, stem, position, scale, and the camera framing, unless the brief asks otherwise. These features keep the subject recognizable while its material changes.

  2. Which parts of the scene must change with the material?

    Suggested answer

    Transparency, highlights, surface texture, and reflections should respond to the ceramic material. A shadow or contact reflection may also need to change if the new material changes the way light interacts with the scene.

  3. What could a model misunderstand in that request?

    Suggested answer

    The model might replace the pear, reshape its stem, or restyle the whole scene. A clearer brief separates the requested material change from the features and regions that must remain stable.

Slide 41 →

Discuss: the ceramic edit

  1. Would you preserve the reflection or let it change? Why?

    Suggested answer

    Let material-dependent reflections change so the ceramic looks believable, while preserving the scene’s lighting direction and spatial relationships. Keeping the original glass reflection exactly could contradict the new material.

  2. When would a mask help more than a longer prompt?

    Suggested answer

    A mask helps when the edit must stay within a specific region or nearby content must be protected. A longer prompt can explain intent but is less precise about pixel location; exact preservation may still require compositing.

  3. What would edges or depth control still leave uncertain?

    Suggested answer

    Edges can constrain outlines and depth can guide spatial structure, but neither fully specifies material, color, lighting, or identity. A plausible shape can still have the wrong surface or lose a distinctive stem.

Slide 44 →

After Part 1: explain the control

  1. How does ControlNet differ from a mask or a text prompt?

    Suggested answer

    ControlNet adds spatial conditioning, such as an edge map, depth map, or pose, to guide generation. A mask specifies an edit region, while text describes the desired content or change; these controls can be combined.

  2. Why can stronger guidance make an edit worse?

    Suggested answer

    If guidance means text classifier-free guidance, increasing it can overemphasize the prompt and produce harsh colors or artifacts. Excessive structural conditioning can instead make the edit too rigid. Stronger control is useful only when it supports the intended change.

  3. What evidence would show that an edit preserved identity?

    Suggested answer

    Compare distinctive features, proportions, and relationships before and after the edit, allowing the requested material change. Use close-up inspection and multiple views where available; overall similarity alone does not establish identity preservation.

Slide 48 →

Before Part 2: artistic intention

  1. What makes an image feel fragile rather than merely attractive?

    Suggested answer

    Fragility might be communicated by thin contours, precarious balance, torn edges, sparse support, or vulnerable spacing. A strong answer identifies a visible feature and explains its effect on the viewer.

  2. Can very different images belong to the same artwork? Why?

    Suggested answer

    Yes. Different images can share a concept, recurring motif, compositional rule, or material language. Explain what connects them and why the differences contribute to the work.

  3. Which artistic decisions would you keep for yourself?

    Suggested answer

    One defensible choice is to retain control of the concept, intended audience response, selection criteria, and final acceptance. Generating variations can be delegated while the artist remains responsible for choosing their purpose.

Slide 69 →

Discuss: two artistic directions

  1. Which poster communicates fragility more clearly? Why?

    Suggested answer

    Either poster can be defended with visual evidence. Torn paper might suggest physical vulnerability, while open space might suggest exposure or absence. Explain which reading best serves AFTER RAIN.

  2. What does the rejected direction reveal about your choice?

    Suggested answer

    The rejected direction helps reveal what the selected version prioritizes. For example, choosing quiet space over busy fragmentation may show that emotional restraint matters more than visible damage.

  3. Which single revision would most change the audience’s reading?

    Suggested answer

    Choose one consequential change, such as reducing the subject’s scale, shifting its balance, or opening a gap in the composition. Predict how that change would alter the viewer’s sense of fragility, then compare versions.

Slide 71 →

A gallery conversation

  1. Which visible decision carries the idea?

    Suggested answer

    Name a specific decision, such as a fragile silhouette against a large empty field, and explain how it communicates the idea. Describing the image as beautiful is less useful than locating the feature that makes it meaningful.

  2. What changed without a clear artistic reason?

    Suggested answer

    An added texture, stronger saturation, or altered lighting may lack a clear reason if it changes the mood without supporting the concept. Identify the change and ask what it contributes before deciding to remove it.

  3. Which revision would most change the audience’s reading?

    Suggested answer

    Prioritize a revision that affects meaning: for example, move the pear closer to an edge to increase tension. State the expected audience response and compare it with the current composition.

Slide 72 →

After Part 2: defend a decision

  1. How can a technically correct image fail as an artwork?

    Suggested answer

    An image can satisfy the material edit, preserve identity, and contain no obvious artifacts yet communicate little about the artist’s intention. Technical correctness supports the work, while artistic judgment asks what the result expresses.

  2. Which visual rule should remain across a series?

    Suggested answer

    Choose a rule connected to the concept, such as the relationship between subject and empty space, a restrained palette, or consistent treatment of torn edges. Keep enough continuity to connect the series while allowing purposeful variation.

  3. What evidence makes a critique useful for the next revision?

    Suggested answer

    Useful critique identifies a visible feature, explains its effect, connects it to the intention, and proposes a testable revision. For example: enlarging the pear reduces its vulnerability; try a smaller version and compare the reading.

Slide 74 →

Before Topic 3: images in motion

  1. What new failures become possible when an image moves?

    Suggested answer

    Motion introduces flicker, identity drift, unstable textures, inconsistent lighting, implausible movement, and incorrect occlusion. A sequence can fail between otherwise convincing still frames.

  2. What should happen when the pear disappears behind a column?

    Suggested answer

    The column should hide the appropriate parts of the pear as it moves behind it. On reappearance, the pear should retain its identity and material, with position and motion consistent with its trajectory.

  3. Can four convincing frames prove that a video works?

    Suggested answer

    No. Four selected frames may miss flicker, a brief identity swap, or an impossible transition. Review the full clip at normal speed, then inspect difficult moments frame by frame.

Slide 103 →

Discuss: the difficult moment

  1. Which moment in this shot would you inspect most closely?

    Suggested answer

    Inspect the approach to the column, partial occlusion, and especially reappearance. These moments test whether the system maintains the subject when much of it is temporarily invisible.

  2. What should remain recognizable after the column?

    Suggested answer

    The pear’s distinctive stem, silhouette, proportions, and intended material should remain recognizable. Pose, visible area, and reflections can change naturally with viewpoint and motion.

  3. What failure would make you reject an attractive clip?

    Suggested answer

    A changed stem, a pop in position, an object visible through the column, or a brief material swap could justify rejection. Judge the clip against the intended event and continuity requirements as well as its attractive frames.

Slide 110 →

After Topic 3: judge the whole shot

  1. What can an approved keyframe establish—and what can it not?

    Suggested answer

    An approved keyframe can establish composition, subject appearance, and the desired material or style at one moment. It cannot establish motion quality, continuity, or correct behavior during occlusion across the clip.

  2. Why might a low frame-difference score reward a bad video?

    Suggested answer

    A frozen or barely moving video can have very small differences between adjacent frames while failing the requested action. Temporal evaluation must check intended motion and events as well as unwanted changes.

  3. Which evidence would convince you that identity survived motion?

    Suggested answer

    Review the full clip and compare distinctive identity features before, during visible portions of, and after a turn or occlusion. Track motion continuity and inspect difficult transitions rather than relying on selected stills.

Slide 120 →

An Alien Mind

  1. How does achieving an assigned goal differ from respecting human values?

    Suggested answer

    A system might achieve a narrow target, such as maximizing visual impact, while disregarding consent, context, or the artist’s intended meaning. Respecting human values requires behavior consistent with those broader constraints, including in unfamiliar situations.

  2. Which claim in the reading is a forecast, and what evidence would you seek?

    Suggested answer

    Choose a forward-looking claim about future capability or alignment and explain why it is a prediction. Relevant evidence could include reproducible evaluations on unfamiliar tasks, independent replication, and documented failures. State what would weaken the claim as well as support it.

  3. What would you keep under human control in an AI-assisted creative project?

    Suggested answer

    A defensible boundary keeps the project’s purpose, consent decisions, representation of people, and final publication under human review. Delegate bounded production steps with clear constraints and inspect the results before acceptance.

Slide 121 →

How to fix your entire life in 1 day

  1. Which idea could help an artist choose what to make, and why?

    Suggested answer

    Examining the goals and habits behind repeated choices could help an artist identify a direction. Turn that reflection into a small project with a stated intention and a concrete criterion for choosing among results.

  2. Which claim seems useful or overstated? What experience supports your view?

    Suggested answer

    One possible response finds a day of structured reflection useful as a starting point but questions a promise of complete lasting change in one day. A fictional or public creative example is enough; sustained changes in practice would provide stronger evidence than an initial feeling of motivation.

  3. How could easy AI generation support a creative intention or distract from it?

    Suggested answer

    AI can support intention by producing a few controlled variations that test one artistic question. It can distract through endless generation and selection by immediate attractiveness. Set the question and stopping rule before generating.

Slide 122 →

Introducing Aleph 2.0 and Edit Studio

  1. What can an edited keyframe establish, and what must the full clip show?

    Suggested answer

    The edited keyframe can establish the intended appearance and composition at one moment. The full clip must show that the edit survives motion, changing visibility, and lighting while preserving the scene elements that should remain stable.

  2. Which demonstration or preservation claim would you test more carefully?

    Suggested answer

    For example, test a preservation claim with a pear that passes behind a column and turns. Check reappearance, stem identity, background stability, and material consistency across repeated outputs; a selected demonstration alone does not establish reliability.

  3. How could this workflow serve AFTER RAIN’s intention rather than dictate it?

    Suggested answer

    Begin with AFTER RAIN’s intended sense of fragility, approve an image that expresses it, and choose motion that develops that idea. Accept the workflow only if the full clip sustains that purpose; technical novelty alone is not a selection criterion.

Slide 123 →

Your final judgment

  1. Where should an artist place the boundary between control and surprise?

    Suggested answer

    Control the constraints central to the work’s meaning, such as identity, composition, or an event. Allow surprise in areas where variation might enrich the concept, then judge unexpected results against the same artistic intention.

  2. How would you judge an edit: technical success, artistic purpose, or both?

    Suggested answer

    Judge both. First check whether the requested edit and continuity requirements are satisfied; then ask whether the image or shot communicates its purpose. A successful result should pass both forms of review.

  3. Which decision would you delegate to AI, and what evidence would you require?

    Suggested answer

    For example, delegate generating material variations while retaining the brief and final selection. Require before-and-after comparisons, checks of protected regions, and full-clip review for video before accepting an output.

LECTURE 02 · THREE PERSPECTIVES

Read, then question.

AI capability, human direction, and video editing. Choose one reading for your Quiz 02 reflection.

Reading 1 · Research perspective and policy argument

An Alien Mind ↗

Pachocki argues that increasing AI capability raises difficult questions about alignment and monitoring. He distinguishes accomplishing an assigned goal from generalizing human values in unfamiliar situations. Read this as a research leader’s position about the direction and governance of AI development, including claims and forecasts that deserve scrutiny.

Reading route & discussion

Read the opening, then focus on “Teaching machines to love” and “Monitoring generalization.” Skim the remaining sections to understand the author’s larger argument.

  • How does the author distinguish goal alignment from value alignment? Explain using an example of your own.
  • Which statement in the article is an observation, which is a forecast, and which is a recommendation? What evidence would you seek for each?
  • If a creative agent achieves a requested visual effect while undermining the artist’s purpose, what should the artist inspect or change?

Discussion connection. Which task could an exhibition team delegate to AI, which artistic decision should remain human, and what evidence would justify acceptance?

Reading 2 · Personal-development essay

How to fix your entire life in 1 day ↗

Koe proposes reflecting on identity and goals, interrupting habitual behavior during the day, and turning the evening’s reflections into concrete actions. Use the essay as a prompt to examine creative habits and intentions. Its promise is a proposition to evaluate, rather than an established guarantee of transformation.

Reading route & discussion

Read the opening and section VI, which presents the day-long reflection process. Follow the argument around those sections as time allows. Completing the full personal exercise is optional.

  • Which part of the proposed process could help an artist decide what to make, rather than only produce more?
  • Identify a claim you find useful and one you would challenge. What experience or evidence would change your view?
  • How could abundant AI-generated options reinforce an unhelpful habit, or support a deliberate experiment? Make the mechanism specific.

Discussion connection. How could a small creative experiment clarify an intention, what observation would make it useful, and where might the essay’s advice fail?

Reading 3 · Public AI product announcement with visual examples

Introducing Aleph 2.0 and Edit Studio ↗

Runway describes a video-editing workflow in which an edited image establishes the desired appearance before the change is applied through a clip. The announcement presents localized edits and preservation as product capabilities. Read the examples alongside the tests you would require for your own shot.

Reading route & discussion

Read the sections describing image-level control and what the workflow makes possible. Inspect one visual example, then separate what the page shows from what it asserts.

  • What decision can the edited reference frame settle, and what remains uncertain until the full clip is reviewed?
  • Choose one preservation claim. What input clip and failure event would make a useful test?
  • How could this workflow support the AFTER RAIN concept without letting a product demonstration determine the artwork’s purpose?

Discussion connection. What should an edited keyframe establish, which motion event is most likely to reveal failure, and what evidence should be reviewed before acceptance?

RESEARCH COAUTHORED BY HARRY YANG

From research
to the studio.

Four collaborative projects connect technical mechanisms to artistic decisions. Published figures and author credits appear in the slides.

LECTURE RESOURCES

Download the complete lecture.

QUIZ 02 · 5 QUESTIONS · 20 POINTS

Connect technique
and intention.

Four questions on lecture content and one reading reflection. Due by semester end.

Open Quiz 02 →