- Published on
Meta Muse Video sample footage exposed, AI generates up to 10 seconds of video per session
- Authors
- Name
- DeepHorus
On August 19, tech media outlet testingcatalog reported that Meta is about to launch an AI video generation model called Muse Video, and has already invited some partners for beta testing. According to the media's tests, Muse Video can generate videos up to 10 seconds in a single run, demonstrating high potential in detail presentation, understanding of scene-object relationships, and temporal stability of continuous frames.
Meta's launch of Muse Video is seen as an important move in the field of generative AI video. Compared with existing AI video tools, the model places greater emphasis on understanding the laws of the physical world, striving to maintain visual coherence in short clips. In the test version, the model can generate cinematic scenes from complex text prompts, such as a woman on a Tokyo street at night and a rocket exploding into confetti.
However, Meta also acknowledges that the current model has obvious shortcomings. Audio-visual synchronization is insufficient, and there is still a gap in the physical accuracy of high-speed motion. The existing test clips are all around 10 seconds, which is not enough to verify the model's stable performance in longer narratives, consistent character persistence, or high-intensity action scenes. These limitations mean that Muse Video still needs time to mature for commercial use.
The multiple test prompts published by testingcatalog showcase the model's capabilities and boundaries. These include a detailed scene of a woman walking on a Tokyo street, and a dialogue video simulating iPhone handheld shooting, where the prompts require realistic shooting imperfections such as motion blur and overexposed sky, all of which the model reproduced well. There are also diverse prompts such as rocket explosions, cat interactions, and wine pouring, demonstrating its ability to handle object dynamics and changes in light and shadow.
In professional scenario tests, Muse Video attempted shots such as orbiting a chef plating dishes, a flickering neon sign, car lights sweeping across a bedroom wall, as well as action scenes like sword fights in a bamboo forest and a brawl in a dim warehouse. These tests verified the model's preliminary understanding of depth-of-field control, lighting changes, and physical collisions. However, there is still slight distortion in the coherence of some complex actions, particularly fast-moving blade clashes and boxing blocks.
The exposure of Muse Video has sparked industry discussions about the boundaries of AI video generation technology. Current mainstream tools mostly focus on generating short videos of 3 to 5 seconds, while Meta has extended single-generation duration to 10 seconds and emphasized temporal stability, offering more possibilities for narrative content creation. However, audio synchronization and physical accuracy remain common challenges for all AI video models, and Meta admits they have not been fully resolved.
Looking ahead, Meta plans to optimize the alignment accuracy of audio and video in future versions and expand long-video generation capabilities. Sources familiar with the matter revealed that the team is exploring the integration of diffusion models with Transformer architectures to improve simulation of complex actions. As the scope of testing expands, Muse Video is expected to find applications in virtual creation, advertising, and other fields. However, before breaking the 10-second limit and overcoming consistency shortcomings, its practical applications will remain limited to short clip generation and concept previews.
