Oh no.

Using Firefly, I got the same type of frustrating results that Olly got (one of which you can see above). Given GenAI’s frequently baffling results, was the time spent carefully crafting prompts and endlessly reiterating on them really time well spent? As of writing this, there are no controls in place to dial in or creatively direct the outputs. Using GenAI can feel like pulling the lever on a slot machine. Without being able to look under the hood to see the reasoning these models are using to create outputs, you can only cross your fingers and hope something good comes out. If you're trying to follow brand guidelines, the traditional workflows are still the most efficient way to go. 

Funnily enough, there is an increasing niche of work for freelancers hired to fix AI generated content. This further demonstrates these results are usually not fit for final delivery, and fixing those results can often take more time than just using traditional methods.

Besides the whole getting-results-at-random thing, we also need to consider the implications of copyright law. Tools like Higgsfield AI allow you to generate different angles, backgrounds, and lighting of a person from uploading a single photo. This would absolutely speed up the workflow on a project like this from Kill 2 Birds, but legally, you would need to have talent sign a contract allowing their likeness to be used to train an AI model (and then to generate potentially limitless images for commercial use). What would be a fair contract for that talent to sign? 

Furthermore, at its core, GenAI needs vast amounts of content in order to train its model. Getting explicit permission from each and every artist, director, etc. would be impossible, so this technology has been trained without permission or payment to the creators. This is a known issue, as products such as Moonvalley’s Marey and Adobe's Firefly are specifically only trained on licensed or public domain content to avoid legal issues. While it’s no secret that both of those models are far outperformed by models trained on copywritten or stolen content (such as Runway or Google's VEO), using these better performing models for commercial purposes is potentially gambling with copyright law. 

Given all of this, it’s not surprising that a recent report from MIT concluded that 95% of companies who heavily invested in GenAI have seen 0% return on investment.

Despite all of those drawbacks and caveats, I do believe there are absolutely uses for GenAI in a motion design workflow. Models like Firefly can be a fantastic partner in creating highly customized static stock assets. Rather than spending hours going down the Getty Images rabbit hole (and possibly finding something that only fits your project 75% of the way), you can prompt Firefly to create a static asset in moments that’s a better fit than any existing stock photo would be. 

Firefly video generation is still a ways away from being the final output machine, but using it for stock assets such as backgrounds like this confetti background for the Dairy Queen announcement I created is extremely efficient. No more time spent begging a producer to buy the license and download the correct size, and no more purchases of $400+ stock assets for a post that is ephemeral in its very nature.

GenAI can also be a good concepting and previs partner. Instead of spending hours pulling visuals that aren't quite right or starting on a brief from scratch, you can use GenAI to help generate ideas and visuals to jumpstart the creative process.

Lastly, these tools are a gamechanger for video editors. Transcribing hours of video into text format would normally take hours in a traditional manual workflow, but now a tool native to recent versions of Premiere can do it in seconds. Trimming a music track to fit a certain length video normally takes a good ear and patience, but now there is a tool that can make a pretty decent cut in moments. I wouldn't use this to bypass a professional audio engineer and use it as final output, but this tool can help you quickly get a sense of the track and see if it's working with your picture. No more excuses to simply fade the audio out (I say as someone completely guilty of this)! There are also GenAI voice over tools that let us quickly test out scripts over picture without the need to hire a professional in the early stages of production, enabling you to iterate more quickly. 

In the end, it’s important to calibrate ourselves and be clear on what is actually possible versus what’s hype/vaporware. GenAI is a tool that’s here to stay. Not every application will be useful, and until generative AI can offer more robust controls over the outputs, I believe best use of GenAI in the motion design workflow will remain mostly in the previs/concepting stage.