CSE 5 DISCUSSION / WEEK 4
Image & Video
Generation
Amisha Prasad
CSE 5 DISCUSSION / WEEK 4
From a prompt to an image
A prompt describes content, composition, and style. A model turns that description into numerical representations.
The model uses patterns learned from images and text to generate a new output.
Training changes model parameters. Generating an image uses those parameters during inference.
CSE 5 DISCUSSION / WEEK 4
Diffusion: learning to remove noise
During training
Add noise to examples. Train the model to predict noise or a cleaner representation.
During generation
Start from noise and repeatedly refine it using the prompt as guidance.
The result
Decode the final representation into an image. Different starting noise can give different results.
A limitation
Plausible appearance does not guarantee correct text, anatomy, or physical details.
CSE 5 DISCUSSION / WEEK 4
Video adds time and consistency
A video model must generate appearance and motion across frames, keeping people and objects consistent.
Longer clips and higher resolution usually require more computation within the same system.
A convincing clip can still contain impossible motion or events that never happened.
CSE 5 DISCUSSION / WEEK 4
The environmental footprint
Hardware
Manufacturing chips and building infrastructure use materials and energy.
Training
Large training runs consume electricity before students use the tool.
Inference
Every generation and retry adds computation during everyday use.
Supporting systems
Cooling, storage, networking, and electricity production add impacts.
CSE 5 DISCUSSION / WEEK 4
One benchmark, with important limits
Task | Mean energy per 1,000 queries |
Text generation | 0.047 kWh |
Image generation | 2.907 kWh |
About 62 times as much energy in this study’s tested tasks.
2024 benchmark. These are not universal costs for today’s tools or videos.
CSE 5 DISCUSSION / WEEK 4
Electricity, carbon, and water differ
Electricity
kWh measures energy use. It does not directly measure emissions.
Carbon
Emissions depend partly on how the electricity is generated.
Water
Cooling and power generation can use water. Location and cooling design matter.
Local effects
The same resource demand can have different consequences in different communities.
CSE 5 DISCUSSION / WEEK 4
Choices that affect the total
More outputs, longer clips, and repeated retries can increase total computation. Efficiency gains can coincide with growing use.
For an early draft, consider a storyboard, an existing licensed image, or a smaller preview before generating final media.
Ask what the estimate includes, which system it measures, and what alternative achieves the same goal.
CSE 5 DISCUSSION / WEEK 4
The dark side of synthetic media
Impersonation
A fake voice or video can make someone appear to say or do something.
Abuse
Nonconsensual intimate imagery and harassment can harm real people.
Deception
Fabricated evidence, fake endorsements, and scams can exploit trust.
Consent
A person’s public photo does not establish permission for every new use.
CSE 5 DISCUSSION / WEEK 4
Case study: Taylor Swift deepfakes
BBC News: “Taylor Swift deepfakes spark calls in Congress for new legislation”
This January 2024 report connects nonconsensual synthetic imagery with calls for stronger protections.
The image can be fabricated while the harm to the person is real.
CSE 5 DISCUSSION / WEEK 4
A realistic image still needs context
Check the original source, publication date, and independent reporting before sharing a consequential claim.
Content Credentials can document aspects of origin and editing history. They cannot prove that a depicted event occurred.
No visible label does not establish that an image is authentic.
CSE 5 DISCUSSION / WEEK 4
AI and the arts: possible uses
Exploration
Try compositions or visual directions before committing to a final piece.
Access
Let people prototype an idea without advanced drawing or editing skills.
Production
Support tasks such as backgrounds, rough storyboards, or revisions.
Human choices
Purpose, selection, editing, and interpretation still shape the work.
CSE 5 DISCUSSION / WEEK 4
Four questions about creative work
Training data
Did creators authorize the use of their work to train the model?
The output
Does the result reproduce protected expression or imitate a person?
Human authorship
What expressive choices did a human contribute?
Credit and payment
Who should receive attribution, control, or compensation?
CSE 5 DISCUSSION / WEEK 4
Sora and Disney: the timeline
December 11, 2025
Disney and OpenAI announced a three-year licensing agreement and a planned $1 billion investment.
The proposed scope
More than 200 characters for fan-created media. The announcement excluded talent likenesses and voices.
April 26, 2026
OpenAI discontinued the Sora web and app experience.
September 24, 2026
OpenAI’s discontinuation notice also listed this date for the API shutdown.
CSE 5 DISCUSSION / WEEK 4
What a licensing deal can settle
A license can authorize specific characters and uses within an agreement’s scope.
Permission from a media company does not automatically cover performers’ identities or every artist’s work.
A major partnership cannot guarantee that a tool will remain available.
CSE 5 DISCUSSION / WEEK 4
Scenario: a student arts festival
A campus club needs a poster and a 10-second promotional clip. It has $100 and one week.
Option A: commission a student artist. Option B: use AI-generated media. Option C: combine paid human design with AI drafts.
The club wants an inviting result, fair treatment of creators, and a manageable environmental footprint.
CSE 5 DISCUSSION / WEEK 4
Your recommendation to the club
1 minute
Choose an approach with a partner.
2 minutes
Explain one benefit and one tradeoff. Address creator permission and resource use.
1 minute
Share one safeguard: credit, consent, a spending choice, or a limit on retries.
CSE 5 DISCUSSION / WEEK 4
A useful test for creative AI
Does the tool help communicate the idea in a way that respects the people and materials involved?
A defensible choice explains permissions, human contributions, resource tradeoffs, and how audiences will interpret the result.
Creative possibilities and responsibilities belong in the same decision.
CSE 5 DISCUSSION / WEEK 4
Your topic suggestions
Scan to share a topic you would like us to explore.
What additional topic or question about AI would you like us to explore in class?