And the World Split in Two: How the US and China Built Two Engineering Schools of Graphic GenAI

In July 2026, the generative AI landscape for graphics and video is no longer a single, unified frontier. It has fractured into two distinct, competing engineering philosophies — one American, one Chinese. This isn't just about different models or datasets; it's about fundamentally different approaches to architecture, scaling, and deployment. The result is a technological schism that is reshaping everything from Hollywood VFX pipelines to Shenzhen's manufacturing design tools.

A recent deep-dive on Habr by the MTS team (see full analysis here) lays out this split in stark detail. The article examines how, over the past three years, the US and China have evolved two separate "engineering schools" for graphical generative AI — schools that now rarely borrow from each other. This isn't a story of one side winning. It's a story of divergence.

The Two Schools: A Tale of Design Philosophies

The core difference, as the Habr piece explains, lies in how each school treats the foundational architecture. The US school, led by companies like OpenAI, Google DeepMind, and Stability AI, has doubled down on the diffusion transformer (DiT) paradigm. These models are massive, often requiring clusters of thousands of H100 or B200 GPUs for training. Think Sora, Veo, or the latest Stable Diffusion iterations. The philosophy here is: train the biggest possible model on the most diverse data, then compress or distill it for deployment.

The Chinese school, represented by players like ByteDance (with its Jimeng model), Kuaishou (Kling), and Alibaba (Tongyi Wanxiang), has taken a different path. They have prioritized efficiency and edge-deployment from day one. Instead of pure DiT scaling, they often employ hybrid architectures — combining diffusion with efficient convolutional backbones or novel latent-space compression techniques. The goal isn't just a beautiful 4K video; it's a beautiful 4K video that can be generated on a mid-range GPU or even a smartphone in seconds.

Why the Split Happened

The Habr article argues this isn't accidental. It's driven by market realities. The US market is heavily oriented toward enterprise and high-end creative tools — film, advertising, game cinematics. A studio can spend $50,000 on compute for a single 30-second ad if the quality is unmatched. The US school optimizes for that ceiling.

China's market, by contrast, is mobile-first and cost-sensitive. Short-video platforms like Douyin (TikTok) and Kuaishou have billions of users who expect instant, cheap, and good-enough generation. A model that requires a $30,000 GPU to run is a non-starter. The Chinese school optimized for the floor — making generation accessible on a $600 phone. This led to innovations like 4-bit quantization of diffusion models and single-step distillation that the US school only began exploring later.

Real-World Case Studies: Kling vs. Sora

A vivid example from the article is the comparison between Kuaishou's Kling and OpenAI's Sora. Both can generate 60-second videos from text. But the engineering paths diverge sharply.

• Sora (US): A pure DiT architecture trained on millions of hours of high-resolution video. It produces stunning, physically plausible scenes — water flowing, leaves rustling — but inference takes minutes on a top-tier GPU. It's a cloud-native beast.
• Kling (China): Uses a 3D Variational Autoencoder combined with a temporal attention mechanism, but with extreme memory optimizations. It can generate a 2-minute video in under 30 seconds on a single A100. The output is slightly less coherent on complex physics but perfectly adequate for short-form content.

The Habr piece notes that Kling's team publicly stated they designed the model to run on a single consumer GPU (RTX 4090) from the start — a constraint that forced elegant engineering solutions. Sora's team, by contrast, never considered consumer hardware as a target.

The Software Stack Divergence

The split extends to the software tooling. The US school has rallied around PyTorch and the Hugging Face ecosystem — open-source but dominated by a few large foundations. The Chinese school, driven by companies like Huawei and Baidu, has invested heavily in MindSpore and PaddlePaddle, their own deep learning frameworks. These frameworks offer lower-level hardware control, which is critical for squeezing performance out of homegrown chips (like Huawei's Ascend series).

This means that a researcher trained on PyTorch will find it difficult to port a model to the Chinese ecosystem, and vice versa. The article describes this as a "software Berlin Wall" — two separate worlds of APIs, debugging tools, and deployment pipelines.

The Geo-Political Angle

It would be naive to ignore the elephant in the room: export controls. The US restrictions on high-end GPU sales to China (like the H100 and B200 bans) have forced Chinese companies to innovate on less powerful hardware. The Habr article explicitly states that this constraint became a catalyst. Chinese researchers couldn't just throw more compute at problems; they had to find algorithmic shortcuts. This led to breakthroughs in sparse attention mechanisms and training-free model merging that the US school, with its abundance of compute, never prioritized.

Conversely, the US school's access to unlimited compute allowed it to explore massive-scale training runs that Chinese teams simply couldn't replicate. The result is a stalemate: the US has better raw quality, but China has better efficiency and deployment scale.

What This Means for Practitioners

For a developer or artist reading this in 2026, the choice of school is no longer just a technical preference — it's an ecosystem lock-in. If you want the highest quality for a feature film, you likely go US. If you want to build a real-time mobile app with generative video effects, the Chinese school offers off-the-shelf solutions that are lighter and faster.

The article also highlights a worrying trend: interoperability is decreasing. Models from one school rarely work well with inference engines from the other. A LoRA trained for Stable Diffusion won't easily port to a Chinese platform. This fragmentation means that the industry may see a future where two separate "app stores" for generative AI emerge — one for each hemisphere.

The Unintended Collaboration

Despite the split, the Habr piece points out one area of convergence: research publishing. Both schools still publish papers in top venues (NeurIPS, CVPR, ICML). The US school leads in theoretical contributions (e.g., scaling laws for video diffusion), while the Chinese school leads in systems and efficiency papers (e.g., training-free acceleration). A savvy researcher reads both. The article notes that the most innovative teams are those that cherry-pick ideas from both schools — using the US school's architectural insights but applying the Chinese school's optimization tricks.

Conclusion: Two Worlds, One Future?

The Habr article concludes with a sobering prediction: this split is likely permanent. As each school deepens its own stack — from chip design to training data curation to deployment frameworks — the cost of switching becomes prohibitive. The world of graphic GenAI has truly cracked in two.

For the rest of us, this means more choice but also more complexity. The golden age of a single, universal generative model is over. We now live in a bipolar AI universe. And whether that leads to faster innovation or a technological cold war remains to be seen.

Source: The analysis in this article is based on the detailed report published on Habr. Read the original for full technical depth: Habr article