EXCLUSIVE: Cascadeur Talks Inbetweening, Ethical AI Training & the Future of the AI vs. Anti-AI Debate

No doubt you've already heard about the recent "video game marketplaces don't need AI disclaimers" discourse over on Twitter, involving Epic CEO Tim Sweeney and many other individuals well-known within the artistic community. What's your take on that?

Eugene: I agree with Tim Sweeney and with Matt Workman, who originally raised this topic. The “Made with AI” label is losing its meaning because most software already uses AI in one way or another, including generative AI. And we will only see more of that. Even Photoshop and Maya are integrating AI tools, and coding is often done with Copilot or Cursor. And on top of that, it simply cannot be verified. So this label can't really serve its intended purpose and ends up being meaningless.

Ultimately, players care about the final quality of the game – its artistic value, the meaning the creators put into it, the enjoyment it delivers. There's always been plenty of low-quality, derivative content even before the AI boom. We called it "generic content." There's just a lot more of it now, and that's a new challenge for moderation.

There's also the separate issue of how ethically training data is collected. But that is a much bigger topic than the question of Steam labels. It's something the entire world, governments and corporations will eventually have to figure out how to regulate. It happens with any new major technological shift.

A lot of people on social media see your software as a prime example of how developers should integrate AI into their tools, what do you think about those reviews?

Eugene: Even artists and users who openly dislike AI tend to speak positively about Cascadeur. The main reasons are pretty simple: there are no concerns about the ethics of our training data, the tool gives animators full control, and AI just assists and removes the routine work. It's fundamentally an AI-assisted approach.

I believe this is exactly the direction the industry will eventually follow. If we accept that AI is going to reshape the job market, then we should expect new tools that improve work efficiency to appear and evolve. And I think those tools will be specialized, AI-assisted systems designed for professional creative people – the kind of tools that help them do more and better, not replace them.

Now, let's talk about the Inbetweening feature. First of all, when did you first come up with it? What practical purpose does it serve? How does it benefit 3D Animators?

Eugene: Initially, we were focusing on physics. It already helped save time and increased the quality and realism of animation. But even in the simplest cases, there was still a huge amount of manual work. Animators had to pose every character's step or create a lot of in-between frames for various transitions. So we decided to experiment using a neural network to generate motion between keyframes, starting primarily with locomotion. It turned out better than we anticipated. This method was far more universal and could produce many different types of movement.

And it really speeds up the process. From just a few poses, you can quickly get a full sketch of the animation and refine it as much as you want, all the way to a completely hand-crafted result. Physics is still there as well, helping you quickly refine trajectories and add interesting details wherever you need them.

The main challenge in animation today is that many types of movement take weeks to produce, and I know from experience that this isn't always the most enjoyable or creative part of the job. I'm confident that animators will be able to work faster – and that's a good thing, because it means their work becomes less routine and more creative.

How does Inbetweening actually work? If I understand correctly, all you need to do is keyframe the poses, and the system will understand what animation it should generate all on itself, but it can't be that simple, can it? How well does it handle more complex sequences with, for instance, erratic character movements?

Eugene: Yes, at the basic level, it really is as simple as that: the tool instantly generates motion between any two poses and, in fact, it can create transitions across any number of poses. It works best with locomotion – walking and running in any direction, but it can also handle jumping, sitting down, lying down, standing up, and even action-style movements if you give it action poses. Some things work better than others, of course, but we're continuing to expand our dataset and train new models.

The workflow itself is very pleasant and naturally goes from broad strokes to fine detail. For example, you set two different poses for your character in different places in space. Instantly, you get an animation where the character moves from one location to the other, matching exactly the poses you set. From there, you can adjust foot placement in the in-between frames, tweak timing, refine the trajectory, etc. You can animate any body part using classic interpolation. And at any moment, you can convert the generated frames into a set of keys and interpolations, and continue working with the animation the way you want. You have full control over the result, but you save a lot of time because you're not creating everything from scratch – you're making adjustments

Inbetweening also helps when creating smooth transitions between different animations. You can also delete any segment of the existing animation and simply regenerate it. You can create many variations of the same animation – for example, a character walking or turning in different directions. These tasks usually aren't the most creative part of the job, and there are typically a lot of them. Now they can be done quickly.

But even if it feels like magic, what happens inside is extremely complex. It took many years for this puzzle to come together. This system includes a sophisticated interpolation engine capable of switching between FK and IK, a retargeting system without which inbetweening wouldn't work across different character rigs, and a very advanced method for detecting and cleaning foot fulcrums, which also suppresses knee-popping. But the core of the tool, of course, is the neural network that generates the animation itself.

The thing many Digital Artists care about the most is how anything with "AI" in its name was trained. So, how was Inbetweening trained? What data have you used?

Eugene: We used Xsens mocap suits and recorded about five hours of motion. Most of it is walking and running in different directions, including run-ups, stops, turns, and so on. We also used animations from Nekki's own games, such as Shadow Fight 3, Shadow Fight 4, and Spine. One of the challenging parts of the process was labeling the data so that the neural network could reliably distinguish between walking, running, jumping, and other actions.

We're continuing to expand the dataset with new types of movement and are training new models for Inbetweening as we speak.

Where do you see all this AI/anti-AI debacle in, let's say, a year? It seems like we've got a typical thesis-antithesis-synthesis situation on our hands that, by its very nature, is bound to resolve itself sometime, somehow – what do you think that synthesis will actually look like when it does? Or maybe that's just wishful thinking on my part, and AI and art will never really coexist, with only a few tools like Cascadeur offering false hope that they can. I'd love to hear your thoughts on it.

Eugene: My personal opinion is that the war against AI can't be won. I don't think today's generators will be enough to reshape the market on their own, but sooner or later, AI will be integrated into production tools and start bringing real value. And once that happens, it will be everywhere.

I don't know how the global community will resolve the copyright and training-data issues for systems that require massive datasets, but at least most specialized tools can be built on the specific datasets collected for them.

And on the optimistic side, I think AI will ultimately help democratize game development and content creation. The entry barrier will decrease, and small teams and companies will be able to create things that previously only large studios and corporations could make. This is already happening today with the development of new tools and the rise of Unreal Engine. AI will simply accelerate that process.

Eugene Dyabin, The Founder of Cascadeur

Interview conducted by Theodore McKenzie