AI
An image of a firefly from Adobe Firefly.
A floppy-eared, brown-eyed beagle turns her head. A sunbeam shines through the driver’s side window. The dog is outfitted in the finest wide-brimmed sun hat, which fits perfectly atop her little head.
If this hat-wearing dog weren’t a clue, I’m describing an AI video. There are other hints too: If you look closely, the dog is sitting snugly between two black-leather seats, which are way too close together. Outside, cornfields and mountains start to blur, and the road contorts behind the car.
Despite these problems, this is still one of the better text-to-video generation models I’ve encountered. And it’s not from a major AI startup, but rather from Adobe, the company behind PhotoShop.
Adobe first released its AI model, Firefly, for image generation in March 2023 and followed it up this month with a video generator, which is still in beta. (You can try out the program for free, but we paid $10 after quickly hitting a limit of how many videos we could generate.)
Firefly’s selling point isn’t just that it makes high-quality video clips or that it integrates with the rest of the Adobe Creative Cloud. Adobe also promises that its AI tools are all extremely copyright-safe. “As part of Adobe’s effort to design Firefly to be commercially safe, we are training our initial commercial Firefly model on Adobe Stock images, openly licensed content, and public domain content where copyright has expired,” the company writes on its website.
In the past, Adobe has also offered to pay the legal bills of any enterprise user of Firefly’s image model that is sued for copyright violations — “as a proof point that we stand behind the commercial safety and readiness of these features,” Adobe’s Claude Alexandre said in 2023. (It’s unclear if any users have taken the company up on the offer.)
eMarketer’s Gadjo Sevilla said that Adobe has a clear selling point amid a fresh crop of video tools such as OpenAI, ByteDance, and Luma: its copyright promises. “Major brands like Dentsu, Gatorade, and Stagwell are already testing Firefly, signaling wider enterprise adoption,” Sevilla said. “Making IP-safe AI available in industry-standard tools can help Firefly, and by extension Adobe, gain widespread adoption in copyright-friendly AI image generation.”
But Adobe’s track record isn’t spotless. The company had a mea culpa last year after AI images from rival Midjourney were found in Firefly’s training set, according to Bloomberg, likely submitted to the Adobe Stock program and slid past content moderation guardrails.
Firefly’s video model is still new, so public testing will bear out how well it’s received and what exactly users get it to spit out. For our trial, we asked for “an extreme close-up of a flower” and selected settings for an aerial shot and an extreme close-up.
We also asked Firefly to show us President Donald Trump waving to a crowd. It wouldn’t show us Trump because of content rules around politics but gave us some other guy.
And, of course, we asked to see if Mickey Mouse — who is at least partly in the public domain — could ride a bicycle. At least on that front, it’s copyright-safe. You’re welcome, Disney.
When compared to OpenAI’s Sora video generator, Firefly takes longer (about 30 seconds vs. 15 for Sora) and is not quite as polished. But if I get into trouble using Adobe’s products, well, at least a quick call to their general counsel’s office should solve my problems.