SenseTime’s AI Infrastructure Platform Enables HiDream.ai to Seamlessly Migrate Video Generation Workloads to Domestic Computing Infrastructure
Enabling Large-Scale Video Generation on Domestic Computing Infrastructure
As domestic computing infrastructure moves toward large-scale commercial deployment, how to bridge the critical gap between “compatibility validation” and “real-world production” has become a new challenge for the industry. Recently, SenseTime’s AI Infrastructure Platform and HiDream.ai used video generation as an entry point to establish a complete path from domestic computing adaptation to large-scale application.
For AI companies whose core offerings are image and video generation models, localization is far more than simply migrating a model from one GPU to another. Video generation models, in particular, have large parameter scales, lengthy inference pipelines, and high computational intensity. Every stage—from underlying computing resources and inference frameworks to model performance, generation quality, and toolchains—can become a bottleneck in the transition to domestic infrastructure.
As a globally leading innovator in multimodal generative artificial intelligence, HiDream.ai is committed to building next-generation native, fully multimodal world models. Powered by its self-developed HiDream series of large models, HiDream.ai continues to expand into commercial marketing, film and television production, content creation, and other scenarios. Its products currently serve more than 50 million professional users and over 40,000 enterprise customers across more than 100 countries and regions worldwide.
As an important AI infrastructure partner of HiDream.ai, SenseTime’s AI Infrastructure Platform has continuously supported the rapid development and iteration of its models through stable and efficient AI infrastructure. This time, the two sides further expanded their collaboration to include domestic computing compatibility validation.
Focusing on HiDream.ai’s core business, SenseTime’s AI Infrastructure Platform started with the replacement of underlying computing resources and worked through key areas including model inference, performance optimization, and generation quality. Through full-stack technical adaptation and FDE expert services, it helped HiDream.ai achieve stable operation in a domestic computing environment, creating a reusable model for the localization of multimodal large-model deployment.
01 Leveraging LightX2V Multi-GPU Parallelism to Achieve a 93% Video Generation Speedup
For visual generation applications, insufficient performance not only increases users’ costs but also affects the product experience and the ability to scale the business. Accordingly, HiDream.ai established clear business targets for this collaboration: efficiently generate high-definition video in a domestic computing environment and optimize end-to-end generation time to an industry-leading level.
Based on these specific objectives, the SenseTime team performed multi-GPU parallel optimization using LightX2V and further explored scalability in multi-node, multi-GPU scenarios. The team first focused on model optimization. Following a joint training-and-inference optimization approach, it significantly improved inference speed while maintaining video quality through training methods such as LightX2V step distillation and CFG distillation.
During inference, the team introduced multiple levels of parallelism, including CFG dual-branch parallelism, Ulysses sequence parallelism, and TP parallelism. These strategies decomposed and coordinated computational tasks throughout the video generation process, improving multi-GPU resource utilization and parallel acceleration while meeting business quality requirements.
Ultimately, when running the DiT model on domestic chips, multi-GPU parallel video generation achieved a 93% speedup, fully unlocking the performance of domestic computing infrastructure.
02 Optimizing Quality to Improve Consistency Across Multi-GPU Generation
If a model generates videos faster after being migrated to a new computing environment, but shows noticeable changes in facial expressions, movements, details, and other features, the migration to domestic infrastructure is still unlikely to enter production successfully.
During multi-GPU inference, the two teams found that the same model could produce subtle differences in output quality between single-GPU and multi-GPU environments, such as variations in facial expressions. To address this issue, the SenseTime team further optimized the multi-GPU communication process. By re-injecting facial features and applying other techniques, the team strengthened feature consistency during multi-GPU inference, bringing multi-GPU output quality closer to that of single-GPU inference.
In another real-world issue, the team found that details such as fingers were unclear near the final frames of some videos. Investigation showed that the issue was related to a mismatch between the number of video frames used during model training and that used during actual inference. The team accordingly supplemented and optimized the inference process, improving detail rendering in long-video generation.
Through these refinements, the SenseTime team addressed, one by one, the engineering challenges encountered when adapting domestic models to domestic computing infrastructure and moving them into real production environments.
03 Supporting Zero-Cost Migration Across Heterogeneous Chips
For HiDream.ai, domestic adaptation involves more than compatibility with a single chip. It must also address the challenge of adapting models across different types of domestic chips.
To solve this problem, SenseTime’s AI Infrastructure Platform developed a unified hardware abstraction system. Through heterogeneous hardware initialization, a unified device abstraction layer, a Registry scheduling factory, hardware operator plug-ins, and other mechanisms, it masks differences between underlying hardware platforms. Based on this unified abstraction system, developers can achieve “develop once, adapt across platforms.” The solution currently supports zero-cost model migration across more than 10 types of heterogeneous chips.
At the same time, SenseTime’s AI Infrastructure Platform completed ecosystem adaptation for mainstream AI development tools such as ComfyUI. Applications can be developed and deployed on domestic computing platforms without changing existing workflows or development habits, substantially reducing migration and development costs.
04 FDE Expert Services: Deep Business Engagement for End-to-End Delivery
For AI innovation companies such as HiDream.ai, model development and product iteration already require substantial resources. Investing significant additional manpower in adaptation and migration within a domestic environment would inevitably increase both labor and time costs.
To address this, SenseTime’s AI Infrastructure Platform brought its engineering expertise directly into HiDream.ai’s business scenarios through its FDE (Forward Deployed Engineer) expert service model.
In this collaboration, SenseTime’s AI Infrastructure Platform provided more than underlying domestic computing resources and platform capabilities. It assembled a professional engineering team to address HiDream.ai’s actual business requirements and participate deeply throughout the domestic adaptation process—from model migration, operator optimization, and multi-GPU parallelism to generation-quality tuning, toolchain adaptation, and final business performance assurance. The team investigated and resolved issues one by one against real business metrics.
This collaborative model, combining “computing infrastructure with engineering expertise,” helps AI companies reduce the substantial engineering investment required for domestic migration, allowing them to focus more of their R&D resources on model innovation and product iteration.
05 Business Deployment: Domestic Computing Infrastructure Powers Large-Scale Short-Video Creation
With the support of SenseTime’s AI infrastructure capabilities and FDE expert team, HiDream.ai’s image models completed compatibility validation on domestic computing infrastructure and now further support large-scale online traffic and short-video creation.
The collaboration between SenseTime’s AI Infrastructure Platform and HiDream.ai has validated a reusable path for supporting video generation businesses with domestic computing infrastructure: starting with heterogeneous computing resources as the foundation; connecting them through a unified hardware abstraction system and development-tool adaptation; focusing on model performance and generation-quality optimization; and then using FDE expert services to engage deeply with on-site business requirements. This ultimately establishes a complete closed loop spanning full-stack adaptation, technical validation, and large-scale production on domestic computing infrastructure.
In the future, the two sides will deepen their collaboration across more AI application scenarios, jointly explore new possibilities for integrating domestic AI infrastructure with generative AI applications, and accelerate the transition of domestic computing infrastructure from “usable” to “easy to use.” Together, they will help AI innovators achieve technological breakthroughs and large-scale business deployment more efficiently.
This article was provided by SenseTime. QbitAI is authorized to reproduce it; all opinions expressed belong to the original author.