Davis Chang：生成式AI與視覺藝術的未來

1. In one sentence

What generative AI actually changed is not whether the picture looks good but who gets to start creating: when production teams compress from hundreds of people to a handful, and when the democratisation of technology lets people with no resources and no connections get a screenplay made, what becomes genuinely scarce next is aesthetic judgement rather than the ability to operate tools.

2. The session

Ø Studio and a project that brought a composer "back to life"

He introduces himself as a co-founder of Ø Studio, having moved from traditional image production into generative AI creation during the pandemic. One project he shared was a collaboration with pianist Ayşedeniz Gökçin: having AI generate new scores in a composer's style, performed by the pianist herself, while AI-generated visuals brought the late composer back on stage as a virtual figure (on stage he named Mozart; on checking public material, this pianist's better-known AI music project involves Chopin, and no independent reporting on this specific Mozart project could be found, so this passage is presented only as his own account from the stage — marked the speaker's account, see section seven). The work subsequently appeared at the Kennedy Center, at TED AI San Francisco and at NASA's Houston Space Center.

First story: a rejected screenplay, made with AI in two weeks

He shared a piece he had wanted to shoot during the pandemic, which the producer and team judged outright to be "simply impossible to make". He later switched to generative AI, adjusting the screenplay and style toward an animated look, and finished the whole film in a very short time to enter a competition run by Runway. He describes that experience as the first time he genuinely realised generative AI brings not just efficiency but "the democratisation of technology": the barrier that used to require convincing investors, studios and distribution platforms in successive gates before anything could be shot is now one that a single person can cross.

How teams got smaller and roles got blurrier

He set the traditional film and television flow (development, pre-production, shooting, post-production) against a generative AI flow: in the traditional flow, "shooting" requires a large, highly divided crew on site, with key personnel potentially scaling from a dozen up to hundreds; in the generative AI flow, "shooting" is almost entirely replaced by "generating", which a handful of people at computers can complete, so the team can compress to three to fifteen. He drew out three trends: teams become small and sharp; job boundaries blur and mix (one person may be both art director and AI tool operator, or a director who is also the cinematographer and the editor); and team composition is configured flexibly per project.

Three tiers of use: previsualisation, theatrical release, social shorts

He splits generative AI's use in film and television into three tiers: previsualisation — quickly producing animatics and lookbooks so a director can communicate with the team or the client; production level for theatrical release and high-quality series — where AI is currently more of a support role, used for effects elements and set extension, with core performance still resting on human actors and quality demands highest; and social media shorts — where AI can already generate quickly and in volume, with style transfer and automated editing and scoring, very efficiently. But he also quoted the preceding speaker Sunny directly, admitting frankly that a great deal of this content is the soulless, endlessly duplicated output of the internet, produced purely for reach.

He also showed several concrete cases: using generation tools to composite a hand-drawn storyboard sketch and a reference image into a complete storyboard with consistent characters and shots; compositing a filmed object and another object into the same clip; and compositing multiple reference images into consistent characters and costumes within one clip.

Hollywood right now: half embracing, half resisting

He shared the industry mood he observes in Los Angeles: Asia is relatively open to using generative AI tools, while part of the Japanese industry is resistant; Los Angeles, with Hollywood's deeply rooted union culture, is split half and half — some worried their jobs will be replaced and resisting, others like him simply taking the screenplays they could never make before and making them. He also observed that production budgets across the industry have been cut continuously since the pandemic, with money flowing more toward streaming platforms and social media, and younger generations more accustomed to shorts and live streams than to going to the cinema; and that even where there is external resistance, the large production companies have long been developing their own AI tools internally. He believes large productions will end up hybrid, because even where effects can be AI-assisted, the fine detail of a human actor's performance (his example being an actor standing there and carrying a scene on a single look) cannot yet be replaced by a machine.

He also mentioned a news item: the director known for *Black Swan* has announced a collaboration with Google to produce AI film (on verification, this corresponds to the collaboration between Google DeepMind and Primordial Soup, the studio founded by Darren Aronofsky — see section seven), showing that some Hollywood directors have formally set off in this direction.

The pace of tool updates: a month is a year

He reminded the room that visual generation models ship new capabilities roughly every two weeks, and that spending thirty minutes a day on community activity and tool updates has become part of his daily practice for staying professional. He described his attitude to rapid iteration as "keep an even mind and go with the flow", rather than clinging to any one tool or technique.

Second story: learning to tell stories from a five-year-old

He shared another experience: a friend's five-year-old daughter can pick up two dolls and say they are married one moment and fighting the next, with the plot shifting continuously and no friction at all — which made him realise how long it had been since he had that kind of creativity himself. So he made the child the "director", directing him in a scene (he lay on the floor pretending to be a patient), then wrote that improvisation into a prompt, uploaded the video and actually produced a generated piece. He also demonstrated taking props lying around at home (imagining a lamp as a machine from *The Matrix*) and having generation tools translate them into a completely different visual style, to make the point that you do not need expensive equipment — imagination alone is enough to start playing.

Character consistency is not that easy

He was equally candid that the "character consistency" features tools advertise take a great deal of trial and error in practice, rather like pulling gacha, and rarely work first time. He demonstrated using a photo of himself plus a reference image with Midjourney's Omni Reference to try putting his own face into different styles (a watercolour, and a mix of Street Fighter and Tekken), and mentioned that one-stop platforms such as Openart, Freepik and Krea now let you switch between different visual models (Kling 2.0, say) in one place.

Conclusion: embrace the uncontrollable, aesthetic judgement is the ticket

He left the room with a question: when everyone can generate a Ghibli-style image, what is creation's real value? He believes the answer lands on aesthetic education — people who know how to break a style down into brushwork, lighting and other details will do very well next, and that is precisely what people with a humanities background and creative experience are good at. He also urged people not to fear mistakes, because generative AI tools are not fully controllable by nature, and rather than chasing perfection it is better to embrace that uncontrollability as a source of surprise. Finally he stressed: even when what you vibe-coded has errors in it, you still have to go and find an engineer, because the engineer is the one with the experience to tell you whether something is any good — so do not stop learning.

3. Figures and cases

• Traditional film and television crews may scale from a dozen people up to hundreds, while a generative AI flow can compress a team to three to fifteen (the speaker's own industry observation).
• Google released Veo 3 in May 2025 and began a collaboration with the AI film studio founded by *Black Swan* director Darren Aronofsky (subsequently verified, see section seven).
• His comparison of AI video tools covered Luma Dream Machine, Runway, Sora, Pika Labs, Google Veo 2, Kling, Minimax Hailuo, Vidu, Wan 2.1, Hunyuan Video and Veo 3, comparing support for text-to-video, image-to-video, first and last frame control, multiple reference images, camera control and motion brushes (this table is drawn from the comparison material he showed on the day).

4. Lines worth keeping

• "Even when what you vibe-coded has errors, you still have to find an engineer, because they're the one with the experience to tell you whether it's any good — so don't stop learning."
• "Imagination is your superpower."
• "A month for humans is a year for GenAI. Stay flexible, keep learning."
• "When everyone can create, what is the real value?"

5. Tools and terms mentioned

• Ø Studio: the AI creative studio he co-founded.
• Runway (Gen-3): the tool he repeatedly mentioned for generating video, style transfer and character compositing; Gen-3 Alpha, released in 2024, leads on text-to-video, image-to-video and motion brushes.
• ComfyUI: the advanced node-based AI workflow tool he mentioned, which allows customised effects and clear tracking of every generation step.
• Pika Labs: the tool he mentioned for compositing objects into existing footage, or compositing multiple reference images into one clip.
• Midjourney Omni Reference: Midjourney's reference image feature for maintaining character consistency.
• Kling: another visual generation model he mentioned, selectable on one-stop platforms.
• Openart / Freepik / Krea: the one-stop AI visual generation platforms he mentioned.
• Google Veo 2 / Veo 3: Google's video generation models; Veo 3 was released in May 2025, adding synchronised dialogue and sound effects.

6. Wider observations

This session is the flip side of 李怡志's session earlier the same day (see "如何生成有效的圖騙"): 李怡志 on how generated imagery gets used to deceive, Davis Chang on how the same technology lets people with no resources and no connections start telling their own stories. Read together they point at the same conclusion — the tool itself has no position, and what decides whether it is deception or a democratisation of creation is the user's judgement and intent. His closing line, "even when what you vibe-coded has errors you still have to find an engineer", also echoes 李怡志's ending in its way: once the technical barrier drops, what people actually need to practise is the judgement to tell good from bad and true from false, not the speed of their hands on a tool.

7. Sources