|
Before we start. Generative AI announcements galore. It might not be obvious at the moment but trust me, 5 years from now you’ll be remembering that historic spring of 2023 when we witnessed one generative AI announcement triggering oohs and aahs after another. The last few days alone there’s been a barrage of groundbreaking generative AI announcements, with more to come from major tech vendors in the next few weeks. I won’t go overboard in announcing them all here, but below's news roundup includes the recent ones that stood out for me. If you’re wondering what this could all mean for your photo & video business, join us at our May 17 Spotlight!
Runway. Text-to-video. First, the hard one: video. Runway has taken some of the wraps of their plans to develop a text-to-video AI application. In case that doesn’t immediately blow you away, think about not just needing to generate a gazillion of frames, but to also align them in the right temporal sequence and add yet another media format, audio, to the equation. Well, we’re not quite there yet. But, Runway is a serious contender: Runway helped develop open-source text-to-image model Stable Diffusion.
Runway currently offers a web-based video editor that specializes in AI tools, such as background removal and pose detection. Its first AI video editing model, Gen-1, was announced in February and focuses on transforming existing video footage by letting users input a rough 3D animation or shaky smartphone clip and apply an AI-generated overlay.
Gen-2, by comparison, will be focused on generating videos from scratch, but, as you’d expect, will initially have lots of limitations.
Adobe. Generative AI tools in beta. And they now have a name: Firefly. Firefly will be a family of generative AI tools. Initial one offers text-to-image features, as well as visual text effects on top of Adobe's fonts. Firefly will be available as a web app for now, but eventually also offered within Adobe’s family of creative apps, along the lines of how Microsoft offers its generative AI tools inside various applications.
Adobe’s model has been trained on images from Adobe Stock as well as images in the public domain or released under an open license. Adobe will also automatically attach credentials to content that was developed with Firefly to indicate that generative AI was used in its creation.
Adobe is also working on developing a universal "Do Not Train" tag for artists who don't want generative AI systems to train on their art. Both tools build on Adobe's Content Authenticity initiative, which we discussed at our most recent Visual 1st conference.
DPReview. The end of an era. After nearly 25 years of operation, DPReview will stop operations on April 10. I'd hate to see it go! A valuable resource for pro, semi-pro and advanced hobbyist photographers worldwide, the site is best known for its in-depth camera reviews. It lost the financial support from its owner, Amazon, which is going through multiple rounds of layoffs and cost cutting exercises.
Midjourney. Getting better and better. Among all the noise of new generative AI tech announcements, it’s easy to ignore that early innovators are moving rapidly to finesse their solutions beyond the stage of “interesting idea but not good enough to use in real life.” Case in point: Midjourney version 5. Check out these examples of Midjourney v5, released last week. More photorealistic, and now even creating the hard part – hands – correctly.
Allcop. Selling AI art. Photo print manufacturer Allcop’s street & pop art site affengeilebilder24.de now offers a collection of “AI Art” that includes the works of well-known AI artists and generative AI imagery created internally by affengeilebilder24.de.
CYME. Cross-application image annotation. CYME launches a new version of its AI-based photo organizer, Peakto. The new update allows photographers to annotate their images across multiple compatible apps through a unified interface. Peakto is compatible with Lightroom, Luminar, Capture, Apple Photos, and, most recently added, Instagram.
Glaze. Beta launch of AI cloaking app. I previously mentioned Glaze, a research project from the University of Chicago that enablea “cloaking” of images in order to make them useless for AI training sets. Glaze now launched a free (non-commercial) beta version of its app for artists to combat the theft of their artistic IP.
US Copyright Office. Images from text prompts are not human authorship. While technology can’t earn copyrights, apparently images created by technology through human-created text prompts can’t be copyrighted either. In a new ruling, the Copyright Office likened text prompts to “instructions to a commissioned artist,” the artist being generative AI applications (i.e. not human and thus not copyrightable).
But images created through technology such as Photoshop are copyrightable. The Copyright Office’s reasoning: ““In the case of works containing AI-generated material, the Office will consider whether the AI contributions are the result of ‘mechanical reproduction’ or instead of an author’s ‘own original mental conception, to which [the author] gave visible form,’ the Office writes in its guidelines. Apparently, the Office doesn’t believe in the concept of “AI prompt artists.” I’m sure this is not the last word on the matter!
ChatGPT-4. Making images supersmart. One aspect easy to ignore among all the excitement about this drastically improved AI chatbot: it now supports multi-modal prompts. In other words: it’s not only capable of creating all its wizardry in response to a text query, but it can now do the same when users submit a photo, or photo with text! Think about it as Google Lens on steroids – and not always steering you towards shopping sites :-). Create a scientific report about whatever that photo signified, give suggestions as who’d be interested in seeing your picture – you get the idea. Or check out the examples Microsoft presented for its multi-modal language model, Kosmos-1. Any photo developer already experimenting with ChatCPT-4? I’d love to hear – drop me a note!
Best,
Hans Hartman
| 
