Orient: Optimistic about the acceleration of AI's penetration into professional film and television creation and the realization of industrial scenarios. Focus on companies with leading multimodal technologies and a complete commercialization loop.

Orient released a research report stating that it is optimistic about the accelerated penetration of AI video in scenarios such as short dramas, advertisements, and professional film production, gradually applying in physical industries like industrialization. The significance of SD2.5 lies in two aspects: first, it pushes film creation to a professional level; second, it starts to expand into industrial field scenarios. The firm believes that attention should be paid to opportunities for vertical multimodal AI applications and is optimistic about technological breakthroughs and cost optimization driving the accelerated development of industry trends, leading to user growth, payment penetration, and commercialization reaching new heights. Furthermore, there is a particular focus on companies that have leading multimodal technology and a commercialized closed loop. Orient's main viewpoints are as follows: MiniMax H3 The firm believes that MiniMax H3 and Seedance2.5 jointly raise the lower and upper limits of the AI video sector, with certain differentiation in model positioning: SD2.5 raises the upper limit, continuing to expand into high-value film tracks and industrial scenarios; H3 lowers the limit, filling existing supply gaps in the competitive landscape with extreme cost-effectiveness and open-source capabilities. The firm is optimistic about the further acceleration of AI video penetration in scenarios such as short dramas, advertisements, and professional film production, and its gradual application in physical industries like industrialization. MiniMax H3: Lowering the limit to fill the supply gap for low to mid-range high cost-performance multimodal solutions. H3's advantage lies in its ability to generate clear 2K output, meeting universal creative demands, while offering a cost-performance ratio of 0.8 yuan/second. The firm anticipates that it could replace a substantial amount of short drama demand previously filled by Seedance2.0; H3 has officially gone open-source as of August 2, allowing short drama companies to independently deploy models and further reduce computing costs. Additionally, H3 demonstrates differentiation in dimensions such as advertisement packaging and visual effects, potentially accelerating the penetration of AI in advertising production. Seedance2.5: Breaking limits, accelerating penetration in professional film production and expanding into industrial scenarios. In contrast to Seedance2.0, which targets general users and the lower market, the update of SD2.5 focuses on enhancing the ability for high-quality professional content creation, further opening up applicable scenarios and customer bases for AI video models. Key iterative capabilities include: (1) generating 30-second direct outputs, raising the previous 15-second limit, allowing segments to fulfill complete narrative arcs and long takes; (2) strengthening all-round reference capabilities, supporting up to 50 materials to avoid the horror valley effect in large-scale generation; (3) improving scene cutting and rhythm control to better fit film creation needs, and enabling video editing and 3D white membrane previews based on timestamps, all emphasizing the enhancement of professional creation limits. Regarding pricing, the API price for SD2.5 at 720p is 42 yuan per million tokens, a 50% increase from 28 yuan at 720p in version 2.0. Although there is a notable price increase, SD2.5 is positioned for professional film creation, where labor and time costs are the primary expenses. Hence, even with price increases, it remains within an acceptable range. In summary, the firm judges that the significance of SD2.5 lies in its ability to push film creation to a professional level and begin expanding into industrial scenarios, such as generating engineering operation guidance videos, constructing intelligent driving training environments, and assisting Siasun Robot & Automation in predicting movement trajectories. The embedding of video models into real production processes is expected to accelerate applications in various physical industries such as industrial manufacturing, embodied intelligence, and smart driving. Risk Warning AI technology iteration may fall short of expectations, AI application landing may not meet expectations, and AI commercialization may underperform expectations.