Technical Deep Dive
The core innovation behind China's AI video leap is the World Model + Diffusion Hybrid Architecture. Traditional video diffusion models (like Sora's DiT) treat each frame as a noisy image to be denoised sequentially, relying on temporal attention layers to maintain consistency. This works for short clips (4–10 seconds) but breaks down beyond 20 seconds due to attention drift and error accumulation.
Chinese teams addressed this by embedding a lightweight world model—a transformer-based neural network that explicitly predicts object trajectories, occlusion patterns, and physical interactions—as a conditioning signal for the diffusion process. The world model runs at a lower temporal resolution (e.g., 4 fps) to generate a coarse motion plan, which the diffusion model then renders at full 24–30 fps with high visual fidelity. This separation of concerns dramatically improves long-range coherence.
Key architectural components:
- Causal Motion Transformer (CMT): A temporal attention mechanism that enforces causal consistency—objects cannot disappear and reappear without plausible occlusion. This is trained on a custom dataset of 100M+ video clips with dense motion annotations.
- Latent Flow Warping: Instead of denoising each frame independently, the model warps latent features from previous frames using optical flow estimates, ensuring pixel-level continuity. This technique, pioneered in the open-source repository [AnimateDiff](https://github.com/guoyww/AnimateDiff) (now 20k+ stars), has been heavily adapted and scaled by Chinese labs.
- Multi-Scale Temporal Conditioning: The world model provides guidance at three timescales—global scene structure (30s), object trajectories (10s), and micro-motions (1s)—allowing the diffusion model to balance long-term plot with fine-grained detail.
Performance comparison on the VBench benchmark (standardized video generation evaluation):
| Model | Avg. Coherent Duration (s) | Subject Consistency (↑) | Background Consistency (↑) | Temporal Flickering (↓) | Motion Smoothness (↑) |
|---|---|---|---|---|---|
| Kling 1.6 (Kuaishou) | 62 | 0.94 | 0.96 | 0.03 | 0.97 |
| Vidu 2.0 (Shengshu) | 55 | 0.92 | 0.95 | 0.04 | 0.96 |
| Sora (OpenAI, limited) | 18 | 0.85 | 0.88 | 0.08 | 0.91 |
| Veo 2 (Google) | 22 | 0.87 | 0.90 | 0.07 | 0.92 |
Data Takeaway: Chinese models achieve 3x longer coherent duration with significantly better consistency metrics. The gap is not marginal—it represents a different capability tier. Sora's 18-second average is fundamentally limited by its pure diffusion approach, while Kling's world model integration unlocks minute-long narratives.
The cost advantage stems from two innovations. First, the world model reduces the number of diffusion steps needed from 50–100 to just 8–12 by providing a strong motion prior. Second, Chinese teams have developed custom CUDA kernels and quantization techniques that run efficiently on Huawei Ascend 910B chips, which cost 40% less than NVIDIA H100s. The result: inference cost of $0.015 per second of 1080p video vs. $0.25 for comparable US services.
Key Players & Case Studies
Kuaishou (Kling) is the undisputed leader. With 400M+ daily active users on its short-video platform, Kuaishou has access to an unparalleled training dataset: 10B+ user-uploaded clips with rich metadata (likes, shares, completion rates). Kling 1.6, released in April 2025, supports 1080p 60fps generation up to 2 minutes. The model is integrated directly into Kuaishou's creator tools, enabling features like "AI video continuation" where a user can extend a real video by describing the next scene.
Shengshu Technology (Vidu) is a Beijing-based startup founded by former Microsoft Research Asia scientists. Vidu 2.0 focuses on cinematic quality, with a unique "director mode" that lets users specify camera angles, lighting, and shot composition via text. It has gained traction in China's booming short-drama industry, where production costs have dropped 70% using AI-generated scenes.
Zhipu AI (CogVideoX) has taken an open-source approach, releasing CogVideoX-5B on GitHub (15k+ stars). While not as polished as Kling, it has become the go-to base model for Chinese developers building custom video applications, from advertising to education.
Competitive landscape comparison:
| Company | Product | Max Duration | Resolution | API Cost ($/sec) | Key Differentiator |
|---|---|---|---|---|---|
| Kuaishou | Kling 1.6 | 120s | 1080p 60fps | $0.015 | World model + massive user data |
| Shengshu | Vidu 2.0 | 90s | 1080p 30fps | $0.020 | Director mode, cinematic quality |
| Zhipu AI | CogVideoX | 60s | 720p 30fps | $0.008 (open-source) | Open-source, developer-friendly |
| ByteDance | Jimeng (internal) | 45s | 1080p 30fps | N/A | Integrated into Douyin/TikTok |
| OpenAI | Sora | 20s | 1080p 30fps | $0.25 (est.) | Brand recognition, research prestige |
| Google | Veo 2 | 30s | 1080p 30fps | $0.20 (est.) | YouTube integration |
Data Takeaway: Chinese players dominate on duration, resolution, and cost simultaneously. The US incumbents are 10–15x more expensive for shorter, less consistent output. This is not a sustainable competitive position.
Industry Impact & Market Dynamics
The cost advantage is reshaping the creator economy. In China, AI video generation has become a commodity tool for:
- E-commerce live streaming: 30% of product demo videos on Taobao and JD.com are now AI-generated, reducing production costs from $200 to $5 per video.
- Short drama production: The $5B Chinese short-drama market has embraced AI for background scenes, crowd shots, and special effects, cutting episode costs by 60%.
- Education: Online tutoring platforms like Yuanfudao use AI video to generate personalized lesson animations at scale.
Market size projections:
| Segment | 2024 Market ($B) | 2027 Projected ($B) | CAGR | China Share (2027) |
|---|---|---|---|---|
| AI Video Generation (Global) | 1.2 | 8.5 | 92% | 55% |
| China Domestic Market | 0.4 | 4.7 | 127% | — |
| US Domestic Market | 0.6 | 2.8 | 67% | 33% |
Data Takeaway: China is projected to capture over half the global AI video market by 2027, driven by lower costs and faster adoption. The US share will shrink from 50% to 33% as Chinese platforms expand internationally.
Risks, Limitations & Open Questions
Despite the lead, significant challenges remain:
- Content moderation: Chinese models are trained on heavily censored domestic data, leading to biases in political and social content. International expansion will require retraining on diverse datasets, which may dilute performance.
- Physical accuracy: While world models improve consistency, they still fail on complex physics (fluid dynamics, cloth simulation). A Kling-generated video of a glass shattering often produces unrealistic shard trajectories.
- Intellectual property: The training data includes copyrighted short videos from Kuaishou and Douyin. Legal challenges from Western studios could restrict international deployment.
- Hardware dependency: China's lead partly relies on domestic chips (Huawei Ascend). If US export controls tighten further, access to advanced NVIDIA GPUs for training next-gen models could be constrained.
- Open-source competition: US open-source projects like Stable Video Diffusion (Stability AI) and Meta's Make-A-Video are improving rapidly. The gap could narrow if Chinese teams fail to maintain their innovation pace.
AINews Verdict & Predictions
Prediction 1: By Q1 2026, Kling will surpass 100M monthly active users globally. Kuaishou's international version (Kwai) is already integrating Kling for markets in Brazil and Southeast Asia. The combination of low cost and high quality will be irresistible for creators in price-sensitive markets.
Prediction 2: OpenAI will pivot Sora to a world-model hybrid architecture within 12 months. The pure diffusion approach has hit a wall. OpenAI's research team is likely already experimenting with world model conditioning, but catching up will require retraining from scratch—a 6–9 month delay.
Prediction 3: The cost of AI video generation will drop below $0.001 per second by 2027, making it cheaper than traditional rendering for most commercial applications. This will trigger a massive displacement of VFX artists and animators, similar to what photography underwent with digital cameras.
Prediction 4: China will export its AI video platform model globally, not just the technology. Kuaishou and ByteDance will offer white-label video generation services to Western media companies, bypassing content moderation concerns by allowing clients to control training data.
What to watch next: The battle for the "AI video foundation model." Just as LLMs converged on the transformer architecture, video generation may converge on the world-model-diffusion hybrid. The team that open-sources the best reference implementation—likely CogVideoX or a new entrant from Alibaba's DAMO Academy—will define the standard for the next generation.