In this interview, we speak with Kfir Aberman, Founding Member at Decart, a company focused on bringing real-time performance to generative video systems. Drawing on his research background at Google and Snap, Kfir shares how Decart is prioritizing latency, personalization, and deployment pragmatism in a field largely centered on offline quality gains. The conversation explores GPU-level optimization, lessons from DreamBooth, product pivots driven by user behavior, and a forward-looking view on interactive AI and AR. Together, these insights provide a grounded understanding of what it takes to transition generative video from demos to live experiences.
Check more interviews like this: Omri Kohl, CEO & Co-Founder of Pyramid Analytics — AI’s Impact on Data Analytics, Decision Intelligence, Citizen Analysts, ROI, Scaling AI, and Emerging Trends
From DreamBooth to Decart, how has your journey through leading research at Snap and Google informed your vision for real-time generative video?
All the major AI players in the industry are now pushing hard toward the development of generative video. At the end of the day, we consume an enormous number of pixels every day, across every digital surface and every form of media. Pixels are the dominant medium of our time, which is why there’s such a massive market and such a race to own the next evolution of pixel generation.
Most companies are moving in the same direction: building bigger, higher-quality, more controllable video models. But the idea that really sparked me wasn’t just quality. It was the realization that while we consume pixels in real time, almost all of this content is created offline, days or months before it reaches us.
If we could generate pixels in real time, if generative video could react to us and adapt to our context, that would represent a true paradigm shift. It would change how content is created, how it’s consumed, and how people participate in it. And while it’s clear the industry will get there eventually, most of the world is still focused on pushing static quality. At Decart, we’ve decided to make real-time the center of our philosophy now, not later.
My time at Google and Snap shaped this mindset. DreamBooth showed me how deeply personalized generative models can connect with people. At Snap, I saw firsthand how powerful dynamic visual experiences can be. Decart is the natural continuation: making these experiences live.
Real-time video generation demands both fidelity and speed. What have been your most impactful strategies for curbing latency without sacrificing quality?
There will always be a fundamental trade-off between speed and quality. That reality doesn’t go away.
But Decart’s advantage is that, unlike most generative-video companies, we didn’t start with generative models. We started as an optimization company. Our founders were mathematicians and PhDs who specialized in optimizing CUDA kernels at a very deep level. They could identify inefficiencies in GPU execution paths that most people don’t even know exist.
This specific optimization layer gave us something rare: the ability to maximize speed while remaining completely lossless in quality. And it’s still the backbone of our competitive edge. If you take our same model weights and run them on a standard GPU server, the model won’t run as fast. On top of that layer, we’ve built a custom stack, bottom-up, that combines CUDA-level optimization, memory-smart execution, compression, and distillation techniques and model-architecture decisions that strike a balance between quality and latency.
This approach is why we’re able to run foundational video models in real time, frame-by-frame, without compromising what people expect from a modern generative system.
Decart emphasizes interactive and personalized AI video. How do you define personalization in this context, and what role does user feedback play?
Personalization is a very broad word. When visual generative AI first gained traction, “personalization” meant generating an image or video of “me,” or of a specific subject based on a reference. That was the DreamBooth era.
But in the real-time context, I define personalization differently.
For me, personalization is about generating content that is unique to me, grounded in my taste, my preferences, and responsive to my reactions in the moment. This is deeper than identity. It’s about behavior, emotional response, visual taste, even the pace or tone at which I like content to unfold.
You’ve co-authored some of the most cited papers in visual AI. Which project posed the most unexpected challenges, and how did it reshape your research approach?
DreamBooth, without a doubt, had the biggest impact and the biggest surprises.
The main challenge was learning how diffusion models actually operate internally and figuring out how to inject new information into them. There were countless possible ways to approach the problem. We ended up choosing a very “expensive” method: full model fine-tuning that took one to two GPU hours just to teach the model what “your dog” looks like.
It wasn’t scalable, but the results were mind-blowing. And once they landed, they unlocked a completely new market. The entire ecosystem exploded with follow-up papers that tried to achieve the same personalization but cheaper and faster.
The lesson for me was profound: Sometimes it’s okay (necessary, even) to start with a heavy, imperfect solution if it introduces an entirely new paradigm.
If it’s valuable, the world will race to make it efficient later.
When building real-time generative systems, how do you balance research innovation with deployment pragmatism?
At Decart, we try to minimize innovation where it’s unnecessary. If a solution exists, even if it’s not glamorous, we implement it and move fast. The problems we pursue are ambitious enough that we’ll hit walls quickly anyway, and those moments force innovation.
So our philosophy is: Move fast where you can, innovate where you must.
At the same time, we recently formed a dedicated team to focus on long-term research bets, big open problems that won’t be solved in a quarter. This gives us a dual rhythm of pragmatic short-term deployment and deep long-term innovation.
What’s a feature or product you’ve envisioned that current hardware or infrastructure still can’t support but will be mainstream within five years?
Real-time perception combined with real-time generation will completely change how we see the world.
Imagine standing on the beach. The system understands your context in real time, but instead of just enhancing the scene, it transforms it based on your preferences. Suddenly, you’re in a “chocolate kingdom” – the sea becomes chocolate, the sand turns into candy, and the entire world becomes a personalized creative layer.
This type of experience is impossible today at the fidelity and speed required for immersion. But with real-time generative models and the next wave of AR glasses, it will feel natural.
I absolutely believe this will be mainstream within five years, and Decart’s technology will play a key role in enabling it.
Could you share a moment at Decart when a product decision required going back to first principles in your research?
When we first developed our real-time video-to-video system, we were targeting the gaming industry. Our vision was to let players apply skins to their games, like playing Minecraft in the style of The Simpsons.
But after launch, something unexpected happened: People weren’t using it for games. They were using it with their cameras. Users wanted to stylize themselves. Their space. Their live video.
This completely changed our trajectory. We had to rethink our data strategy, shift toward human-centric effects, and revisit some of the assumptions underlying our research. It was a classic first-principles moment. The product told us where the real value was.
If you had to describe the soul of Decart’s mission in one word, what would it be, and why?
Reach.
At the end of the day, our goal is to develop new capabilities that will spark new user behaviors, unlock new markets, and reach hundreds of millions or even billions of people.
We view real-time generative video as a medium that can reshape how people express themselves and how the world visually responds to them.
