It is reported that "Luma AI", an American Silicon Valley company in the field of AI vision, recently completed a new round of financing, amounting to US$90 million.

The investment lineup in this round not only includes four European and American companies or funds: Amazon, AMD, Factorial Funds, and LDV Capital, but also introduces the famous Korean consortium, Hanwha Group. At the same time, old shareholders A16Z, Amplify Partners and Matrix Partners continued to increase their investment.

It is understood that this round of financing is mainly used to accelerate the development of basic models and products for visual artificial intelligence.

Founded in 2021, Luma AI is a technology company focusing on computer vision content. Its self-developed models cover video generation, 3D generation and image generation. In January 2024, "Intelligent Emergence" reported that Luma AI completed a US$43 million Series B financing, and the investor was A16Z.

Globally, resource allocation in the AI ​​​​track has entered the "midfield". According to statistics from technology media Techcrunch, the average monthly number of financings exceeding 100 million in the second half of 2024 is 10% less than in the first half of the year. At the same time, hot money is pouring into the AI ​​application layer, especially in areas such as AI search, AI sales, robots, and AI programming.

The model layer is infrastructure, and the AI ​​model layer cannot become a product alone. The final traffic needs to be handled by AI applications - both investors and AI practitioners have reached this consensus.

On November 26, 2024, Luma AI, which mainly works at the model layer, also released the Dream Machine AI creative platform, its first AI application product after the video generation model Dream Machine came out of the industry.

“Compared with language models such as ChatGPT, video models are still a relatively niche field.” Luma AI product designer Jiacheng Yang found that Dream Machine users are mainly professionals with AI or film and television production experience . He explained to "Intelligent Emergence" the reason for releasing an AI creative platform focusing on image design:

"Compared with video generation, the user base in the image field is larger, which is conducive to expanding our user base. We The goal is to make an AI visual tool that can be easily used by both AI novices and design novices. ”

Dream Machine AI creative platform can be understood as a collection of Wensheng image design, AI brainstorming, A design platform with main body/style reference, design drawing to video conversion and other functions.

△The subject/style reference function of Dream Machine AI creative platform. Image source: Luma AI

Compared to Midjourney and StableDiffusion and other text-based image products, the Dream Machine AI creative platform has a stronger ability to understand natural language prompts, and can also generate higher-definition and design-rich text in pictures.

△High-definition text generated by Dream Machine AI creative platform. Image source: Luma AI

The reason why the Dream Machine AI creative platform is easy to use and has strong performance lies in the underlying model capabilities. At present, the platform's language understanding capability comes from Luma AI's Agent built based on a third-party language model; its image generation capability comes from Luma AI's self-developed image generation model Luma Photon; and Tusheng's video capability comes from June 16, 2024 Dream Machine, a self-developed video generation model released today.

At that time, video generation models such as Sora and Vidu were only in the demo stage and were not publicly tested. Dream Machine once became popular on social platforms due to its first "free" and "public beta", as well as its good performance and "meme" gameplay.

Four days after its launch, the number of Dream Machine users exceeded 1 million. At the same time, Barkley Dai, head of data products at Luma AI, told "Intelligent Emergence" that Dream Machine's promotion fee is 0.

Currently, the Luma AI team size is about 50 people. According to Barkley, after deciding to launch the video generation project in December 2023, the team size expanded from 10 to 50 people, mainly introducing top talents in the field of video generation.

The effect of high talent density operations is reflected in the performance of Dream Machine. Dream Machine can currently generate a 5-second video in about 20 seconds. At the same time, the extremely simulated camera movement trajectory, natural light and shadow changes, and rich camera positions are the characteristics of Dream Machine. In version 1.6, released in September 2024, users only need to enter the text prompt to adjust the camera's movement direction.

At the same time, Luma AI, which started out with 3D generation technology, also owns the Text to 3D tool Genie. At that time, Genie was the only tool on the market that could generate 3D models in 10 seconds.

At the commercialization level, on the one hand, Luma AI’s model products in the video, image, and 3D fields provide APIs to the outside world; on the other hand, application layer products such as the Dream Machine AI creative platform will use limited APIs. Free + paid subscription charging model.

At present, Luma AI has become one of the few AI start-ups with comprehensive layout in the fields of video, image, and 3D multi-modality. In public interviews, LJiaming Song, chief scientist of uma AI, mentioned that the amount of tokens required for multi-modal model training is much greater than language, and the multi-modal Scaling Law can allow the model to better understand the world.