AI Engineer World's Fair 2025

The State of Generative Media Today

Loading slides…

Read the talk

The State of Generative Media Today

Gorkem Yurtseven traces the market for generated images, audio and video from fal.ai’s perspective, connecting broader model access with personalized advertising, visual shopping and interactive video.

From a talk by Gorkem Yurtseven

When extraordinary images became accessible

In 2022, Gorkem Yurtseven was sitting on the floor at home, watching Sam Altman reply to people’s requests with DALL·E 2 images. The pictures seemed extraordinary. Even with experience in the industry, he thought OpenAI had established a lead that would be very difficult for anyone else to close. Looking back, those same images already seemed low quality by the time of this talk. His perspective comes from fal.ai, a generative media platform whose inference engine serves image, audio and video models, alongside partnerships with closed-source providers.

Slide displaying a Sam Altman post with an image of teddy bears at computers on a lunar surface, with the speaker inset at lower left.Open full source frame
A DALL·E 2 example shows teddy bears using computers on the Moon.

There had been earlier waves of excitement: GANs, Google’s DeepDream, and consumer applications such as Prisma, which transformed uploaded selfies into stylized portraits. Their applications were narrower, but the ambition to make art with computers was much older still. Yurtseven shows a recreation of Harold Cohen’s work, with a large machine drawing on a huge canvas. Computer graphics and generative graphics belong to that longer history of using computation to produce visual art.

The apparently unassailable lead in image generation did not last. Yurtseven’s release sequence begins with DALL·E 2 on April 6, 2022, followed by Midjourney through a Discord bot and the open release of Stable Diffusion. That last step changed who could participate: people could run comparable technology on home GPUs and build services around it. SDXL followed, then a growing mix of open and closed models, including FLUX, released the summer before the talk. Access expanded from watching someone else generate images to operating the models and building products around them.

0:220:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:22 · section reference included

Cheaper creation changes advertising’s economics

The marginal cost of creation can fall without eliminating the value of creativity. Yurtseven deliberately separates the two. Storytelling and creative decisions still matter; once those decisions are in place, producing the next artifact becomes cheaper. He expects that shift to affect social media, advertising, marketing, fashion, film, gaming and e-commerce, eventually reaching content of every kind.

Advertising offers a precedent for software expanding a media market. Yurtseven claims that YouTube’s advertising revenue exceeds the revenue of other media companies except Disney. He points out that Disney’s total also includes parks, cruise ships and other non-media businesses. The comparison motivates his expectation that another technological change could increase advertising volume, although it is not presented with a consistent accounting basis across companies.

Yurtseven says advertising grew threefold since 2000, with all that growth coming from online or software advertising. The talk does not specify the geography, end year or inflation basis for that estimate. His forecast follows the same pattern: advertising will be among the first industries affected by generative media at scale, and AI will supply much of its future growth.

4:444:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:44 · section reference included

From ad variants to audience participation

Lower production costs allow personalization at several levels:

  • Demographic variants: Generate versions of the same ad for a hypothetical 10,000 demographic groups.
  • Individual context: Generate an ad on arrival, using context such as the website a visitor came from.
  • Interaction: Let the audience’s actions influence what gets generated, making the experience more than a fixed creative asset.

These approaches change both the number of ads a campaign can produce and when production happens: some content can be generated in response to the person viewing it.

Dark slide reading “Ad industry will multiply in size, ads will become hyper personalized and interactive,” with the final two concepts highlighted in purple.Open full source frame
The slide predicts growth in advertising through hyper personalized and interactive ads.

Advertising can also absorb a volume of content that other media cannot. Adding a hypothetical 1,000 movies does not give a viewer more time to watch them. Ads occupy recurring opportunities on phones and television, where a different creative can appear on each encounter. Some consistency still matters, but repeated ad impressions create room for far more variation than a person’s finite movie-watching schedule.

Yurtseven describes a campaign fal.ai worked on the previous year for A24’s Civil War, a film about an imagined civil war in the United States. Its little green toy soldiers became a personalized audience experience:

  1. A participant submitted a selfie and a description through a live marketing website.
  2. The system generated a little green toy soldier incorporating the participant’s face.
  3. Participants’ personalized soldiers appeared on a display in Times Square.

The useful mechanism is the connection between an audience input, a generated representation and a public display. Instead of merely selecting which ad to show, the campaign let people become part of the creative itself.

7:147:32
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:14 · section reference included

Virtual try-on meets an existing shopping need

E-commerce provides another route into an existing market. Yurtseven describes it as gaining roughly “1%” of US retail each year, independently of AI. His wording concerns retail share but does not distinguish percentage-point gains from percentage growth. The relevant direction is the continuing movement of shopping online, where a highly visual experience gives generative media opportunities to add interaction.

Virtual try-on is his clearest example. By the time of the talk, he had seen it develop over several months to perhaps a year into one of generative media’s strongest early examples of product-market fit. Retailers were adopting it, and startups were building around it. Rather than asking consumers to find a use for a new model, try-on puts generated imagery inside an activity they already perform: evaluating something they might buy. That is why he sees every retailer and e-commerce website as a potential generative media user.

9:329:47
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:32 · section reference included

Video repeats the breakthrough—and the catch-up

When Sora appeared, Yurtseven recognized the pattern from DALL·E 2. The February 2024 research reveal looked, if anything, even more impressive. This time, however, he did not interpret OpenAI’s achievement as a durable barrier to everyone else. Having watched competitors catch up in images, he saw a demonstration that comparable video capabilities could emerge elsewhere too.

Yurtseven reports that video was barely represented on fal.ai in October, reached 18% in February and was around 30% when he checked the day before the talk. He introduces these figures as a revenue snapshot but also describes model usage, so the denominator and the October year remain unspecified. He treats fal.ai’s trajectory as a proxy for the wider market: video was growing quickly even while generation remained expensive and results imperfect. From that observation, he predicts video will eventually dominate generative media.

10:4911:02
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:49 · section reference included

Why video could become a much larger market

Yurtseven’s market-sizing argument combines the cost of producing video with an assumption about its usefulness and engagement. He explicitly presents it as rough math:

FactorEstimate or assumption
ComputeVideo models require roughly 20× more compute.
EngagementAssume video is 5× more engaging.
Industry reachVideo serves more industries and use cases.
Market forecastGenerative video becomes 100–250× the size of image generation.

The compute estimate has no specified models, hardware, resolution, clip duration or workload, and the engagement multiplier is hypothetical. Multiplying the two stated numerical assumptions gives the lower end of his forecast; the talk does not provide a numerical derivation for the upper end. These are expectations about market scale, not measured model performance or realized market size.

That forecast does not require image generation to stop growing. Yurtseven expects substantial image-market growth over the following years, while predicting that video will grow faster and ultimately become much larger. The distinction is between relative market sizes, not between a growing medium and an obsolete one.

12:3012:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:30 · section reference included

Sound, then the possibility of real-time generation

The next source of growth is capability, not simply volume. Google DeepMind’s Veo 3 illustrates the progression: first video models improved consistency, then they added sound. Each addition can make a previously impractical use case possible. At the time of the talk, Yurtseven said Veo 3 was not yet on fal.ai; he was looking ahead to what people might build with it in advertising and e-commerce.

Beyond richer clips lies a different operating target: generate one second of video in one second. Yurtseven predicts that faster, cheaper generation will eventually reach that rate, allowing generated content to stream to a user. The change matters because the output could become an ongoing experience rather than a clip produced before playback.

If generated video can respond during an experience, the boundary between a movie and a game becomes less distinct. Yurtseven poses social apps and live events as open possibilities. Fortnite already hosts live events; more lifelike generated environments might extend that kind of participation to parents and other people who do not normally play video games. The question is not only how realistic the imagery becomes, but who can participate and how they interact with it.

13:2713:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:27 · section reference included

Image editing opens another path into enterprises

Image models are still adding capabilities too. FLUX.1 Kontext and GPT-4o image generation bring new editing capabilities and better text rendering. Those improvements challenge the idea that image generation had reached a plateau: a model that can revise an image or render its text more effectively can address work that a visually impressive first draft cannot.

Yurtseven associates these capability shifts with adoption by more mature businesses. As editing and text rendering improve, he expects generative images to fit more enterprise use cases. He closes by inviting people to help build that infrastructure: fal.ai is recruiting machine learning, inference and product engineers, among other roles, and he offers to discuss generative media with attendees during the rest of the day.

15:3215:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:32 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…