Bria’s New AI Learns from Time, Paving the Way for Automated Media

TEL AVIV, Israel – June 10, 2026 – In the relentless push to automate content creation, the biggest hurdles are often the smallest details: the flickering edge of a subject, a halo of light from a poorly removed background, the subtle inconsistencies that scream "this was done by a machine." Today, AI infrastructure company Bria announced a significant step toward solving this problem with the launch of V-RMBG 3.0, a video background removal model that does more than just see frames—it understands time.

This isn't merely an incremental software update. Built on a new autoregressive architecture, the model represents a fundamental shift away from the frame-by-frame processing that has long defined, and limited, automated video editing. By learning from preceding frames, Bria's new tool introduces a form of short-term memory to the process, creating a level of temporal consistency that has, until now, been the exclusive domain of painstaking manual editing. It’s a development that provides a clearer signal of where the entire media production industry is headed: toward fully autonomous, end-to-end pipelines that just work.

The Technical Leap: Teaching AI to Remember

For years, the core challenge in automated video background removal has been its inherent amnesia. Most models treat a video not as a continuous flow of motion, but as a rapid-fire slideshow of thousands of individual images. They analyze a frame, isolate the subject, and discard the background before moving to the next frame to repeat the process from scratch. This stateless approach is the source of the ubiquitous flicker and "boiling" edges that plague AI-generated video composites, as the model second-guesses its own decisions from one moment to the next.

Bria's V-RMBG 3.0 tackles this problem at an architectural level. Its autoregressive design makes it "time-aware." Instead of analyzing each frame in a vacuum, the model uses the output of the previous frame—specifically, the masked subject and its alpha channel—as context for processing the current one. It learns what the scene looks like over time, building a consistent understanding of the foreground subject. The result is a dramatic reduction in artifacts, yielding stable masks and clean edges that hold up even with complex motion, fine details like hair, and semi-transparent objects.

"The next wave of AI-native media isn't being built by humans reviewing individual frames, but by pipelines that run end-to-end without human correction loops," said Dr. Yair Adato, CEO of Bria. "V-RMBG 3.0 is our contribution to that shift: a model that understands time the way a human editor does, so the teams building tomorrow's content infrastructure don't have to work around it." This move from a stateless to a stateful model is a crucial step in building AI that can be trusted with professional-grade output without constant human oversight.

From Lab Bench to Production Pipeline

While the technical innovation is significant, its translation into tangible business value is what will drive adoption. In a market for background removal projected to grow at a 15% CAGR through 2033, efficiency and quality are the currencies of success. Here, Bria’s internal benchmarks suggest a formidable advantage.

In head-to-head tests against major open-source and commercial competitors like MatAnyone 2, Cutout, and Veed, V-RMBG 3.0 was preferred for its output quality over 63% of the time. This validation was not a simple test but a rigorous process involving over 350 diverse video clips, from standard talking-head interviews to dynamic full-body product demos. More striking is the claim of speed: the new model is reportedly up to nine times faster than every competitor tested.

For businesses operating high-volume content pipelines—such as avatar platforms, live commerce studios, and automated video editors—this combination of speed and quality is transformative. The throughput difference isn't marginal; as Bria notes, it's the difference between a tool that fits a workflow and one that bottlenecks it. The ability to process video in real-time or near-real-time without sacrificing quality opens up new possibilities for live streaming and interactive applications, where latency is a critical barrier.

This focus on seamless integration is further underscored by the deployment strategy. "This is the right architectural fix for the problem, and because it ships as a drop-in upgrade, every customer on V2 gets better output the moment they make their next API call," explained Yael Lubratzki Kurman, VP Product at Bria. "No migration, no integration work, no trade-off." This frictionless upgrade path removes a major barrier to adoption, allowing the company's existing user base to immediately benefit from the new architecture.

The Infrastructure for AI-Native Media

V-RMBG 3.0 is more than a product; it’s a foundational piece of Bria’s broader vision to build the infrastructure for "AI-native media." This vision is predicated on creating tools that are not only powerful but also controllable, rights-clear, and deployable within the complex constraints of modern enterprises. The new model’s flexible deployment options are a testament to this strategy.

Customers can run the model in the cloud, within their own cloud environment (BYOC), on-premises, or even on-device. This flexibility is a direct response to the growing enterprise demand for data sovereignty and security. By enabling on-premise and BYOC deployments, Bria allows organizations in regulated industries to maintain full control over their data, satisfying compliance requirements that often preclude the use of standard cloud-based APIs. The on-device option goes a step further, enabling real-time, low-latency applications while ensuring sensitive content never leaves the user's machine—a critical feature for privacy-focused use cases.

Furthermore, Bria is making model weights and inference code available to customers for self-hosted deployment. This level of transparency is rare in the world of proprietary AI and allows for the deep evaluation and auditing that many enterprises now require. It reflects a mature understanding that for AI to become truly foundational infrastructure, it cannot remain a black box.

By solving the temporal consistency problem, Bria is removing a significant roadblock on the path to fully automated content creation. The ability to reliably and quickly separate subjects from their backgrounds is a cornerstone of virtual production, e-commerce visualization, and the booming creator economy. V-RMBG 3.0 demonstrates that the key to building the future of media isn't just about more processing power, but about smarter architectures that allow machines to perceive the world in a more fundamentally human-like way.