Runway’s new 4.5-generation text video model beats OpenAI’s Sora 2, Google’s Veo 3 | Technology News

AI Video & Visuals


New York-based AI company Runway has introduced its latest video generation model, Gen-4.5, which it claims has outperformed competing text-to-video models from OpenAI and Google.

“Runway Gen-4.5 is the world’s highest-rated video model, offering unprecedented visual fidelity and creative control. It produces cinematic, highly realistic output while offering unlimited creative freedom and precise control over every aspect of production,” the company said in its official blog announcing the launch.

The company gained attention after launching its Gen-1 model in 2023. Runway claims that Gen-4.5 will push the frontier of video generation to new heights. The new model shows significant advances in both pre-training data efficiency and post-training techniques.

Additionally, the company said the new model sets new standards for temporal consistency and controllable action generation across different generation modes. In the benchmark named Elo, Gen-4.5 secured a score of 1247 points, slightly ahead of Google’s Veo 3’s score of 1226 and OpenAI’s Sora 2 Pro’s score of 1206 points.

According to reports, this model delivers superior quality without compromising on performance and maintains the speed and efficiency of the 4th generation. One notable aspect of this model is that it offers “unprecedented physical precision” and visual accuracy. This means you can produce videos with realistically moving objects, precise dynamics of flowing liquids, and details like hair and material weaves that remain consistent over time and movement.

When it comes to aesthetics, Gen-4.5 can handle a wide range of styles, from cinematic to photorealistic animation. The Gen-4.5 model is built in collaboration with Nvidia, and its training and inference are performed on Blackwell and Hopper GPUs. Despite the advances, there are reportedly common problems with this model. Runway argued that Gen-4.5 struggles with causality and object permanence.

Additionally, this model is prone to “success bias” which can lead to erroneous results, such as a poorly aimed arrow hitting a target when it should have failed. However, the company believes these issues are critical to building a reliable global model and will continue to iterate on them. This model is available across all subscription plans, giving creators and businesses access at scale.

© IE Online Media Services Pvt Ltd





Source link