SenseTime and HiDream.ai run video-generation workloads on domestic chips at scale

Moving domestic compute from "adaptation verification" to real production is the industry's new challenge, and SenseTime's big-device platform and HiDream.ai (智象未来) have tackled it through video generation, running the full chain from chip adaptation to scaled deployment. For AI companies built around image and video generation models, localization is not a simple GPU-to-GPU swap. Video models have large parameter counts, long inference chains and heavy compute demands, so every link — from base compute and inference frameworks to model performance, output quality and toolchains — can become a bottleneck. HiDream.ai (智象未来) is a global multimodal generative-AI company developing next-generation native full-modality world models with its self-developed HiDream series. Its products cover more than 100 countries and regions, serving over 50 million professional users and more than 40,000 enterprise customers across marketing, film production and content creation. As HiDream's key AI-infrastructure partner, SenseTime started from compute substitution and worked through inference, performance optimization and generation quality, using full-stack adaptation and FDE (Forward Deployed Engineer) expert services to keep HiDream running smoothly on domestic chips. On performance, the SenseTime team optimized multi-card parallelism on LightX2V — combining step distillation, CFG distillation and strategies like CFG dual-branch parallelism, Ulysses sequence parallelism and tensor parallelism — to reach a 93% multi-card video-generation speedup for DiT models on domestic chips. On quality, the teams tightened feature consistency across multi-card inference, re-injecting Face features so multi-card output aligns with single-card results, and fixed detail degradation such as blurred fingers in long-video tail frames. On migration, SenseTime built a unified hardware abstraction layer that hides differences between underlying chips, enabling "develop once, deploy anywhere" with zero-cost model migration across more than 10 heterogeneous chip types, plus adaptation of mainstream tools like ComfyUI so workflows and habits stay unchanged. The FDE expert service embeds engineering capacity directly into HiDream's business, resolving issues item by item against real business metrics. As a result, HiDream's image models have completed domestic-compute adaptation and now support large-scale online traffic and short-video creation. Both sides say the collaboration validates a reusable path for domestic compute in video generation, and plan to extend it to more AI application scenarios, pushing domestic compute from "usable" toward "great to use."