The launch of generative AI video arrived with promises of the ability to create videos in seconds – but a creative director working with a number of exclusively AI-made products suggests that the reality isn’t quite so simple.
During a presentation at Adobe Max Japan, Shimizu Shota, a generative AI engineer, outlined the advantages and disadvantages of using generative AI and, in doing so, disputed the myth that generative AI is “easy.”
Adobe Max Japan coincided with the launch of the Adobe Firefly generative video model earlier this month. Shimizu Shota was among the long list of speakers and, according to the Japanese website Digital Camera Watch, spoke on the advantages and disadvantages of using generative AI for video.
What struck me among Shota’s translated statements were his comments on the misconception that AI makes it easy to create. During the presentation, he walked through what was actually needed to create some of the AI videos produced by DO/AI, an AI-focused agency where Shota works as the creative director.
While some AI platforms advertise generating video in just a few seconds, getting to the final result in fact takes days. In the first video Shota presented, he said it took two days and more than a dozen different AI platforms to produce, from ChatGPT for the prompts to Adobe Firefly and Midjourney for images.
But the second example the creative director shared offers even more insight. The AI short called Sin-chan and Dad is about a minute and a half long. Just one scene in the short required generating more than 600 images, while another required 800 still images. Shota explained that the AI created unexpected results, including changing the character’s clothing mid-shot, so the clip had to be generated as a series of still images instead.
ABOVE: Watch the AI-intensive Shin-chan and Dad
The presentation adds to the ongoing conversation on what generative AI can and cannot do. For example, asking a generative video platform to create a video of a gymnast usually produces rather horrifying results with a clear misunderstanding of complex human movement.
Shota also noted that AI generators can be inconsistent, such as details changing mid-scene. He also noted AI "hallucinations," or generations of things that are physically impossible.
Generative AI video is in its infancy, however, and Adobe’s newly announced Firefly Video model is still in public beta testing. AI's weaknesses aren't necessarily a bad thing – some of the weaknesses of artificial intelligence can be used to spot whether or not an image is real or is made by AI.
You may also like…
Read more on the best video editing software and the best Photoshop alternatives.
