Last time, we compared the most popular AI video generators in the western market — Sora 2, Veo 3, and Kling AI. But because Sora 2 and Veo 3 are not officially available in Hong Kong without a VPN, many readers asked for a similar showdown featuring Chinese tools. So we ran the test again. This round, we picked two of the most widely used AI video generators in China — Alibaba’s Wan 2.5 and ByteDance’s Jimeng AI (即夢) — and put them head-to-head against Sora 2.
What are Sora 2, Wan 2.5, and Jimeng AI (即夢)?
Before we get to the results, here’s a quick look at each tool.
Sora 2 — OpenAI’s 2025 video generation model known for its strong physics, long coherent scenes, and cinematic storytelling. It excels at producing natural, realistic motion and visually polished sequences.
Wan 2.5 — Alibaba’s upgraded model launched in September 2025. It focuses on high-resolution output, accurate motion, and enterprise-friendly workflows through Alibaba Cloud.
Jimeng AI (即夢) — Released by ByteDance in mid-2025, targets short-form content creators. It generates quick, visually appealing clips with a strong emphasis on human characters — ideal for platforms like Douyin.
The Comparisons
As with our previous test, we created three prompts with different levels of complexity and asked all three tools — Sora 2, Wan 2.5, and Jimeng AI — to generate videos using the exact same Chinese instructions. This allowed us to compare each tool’s strengths and weaknesses under equal conditions.
Prompt 1
一名家庭主婦在廚房切檸檬
This simple prompt lets each model interpret the scene freely. Cutting fruit is also a reliable way to evaluate how well an AI understands physics.
We expected strong results from Wan 2.5 because it is the newest model, but its output was surprisingly weak. Not only did the lemon slice in an unrealistic way, the housewife’s appearance looked dated — similar to AI video quality from 2024.
Jimeng performed better in this round. The lemon cutting still didn’t feel fully natural, but the human character looked more realistic than we anticipated. The skin and hair had imperfections, which made her appear more lifelike. On a mobile screen, casual viewers might even believe she’s a real person.
Sora 2 continued to dominate. The movement, physics, and human realism were all convincing. On top of that, it added a small story element by turning the lemon cutting into a scene about preparing sparkling lemon water. This extra layer of creativity is where Sora 2 clearly pulls ahead.
Ranking: Sora 2 > Jimeng AI > Wan 2.5
Prompt 2
特寫鏡頭，金屬質感的咖啡機在暖調環境光下運作。沖煮頭緩緩落下，細緻的蒸氣飄散。深色的濃縮咖啡從金色濾嘴穩定流出，形成絲滑的深褐色拉絲，落入白色陶瓷杯中。鏡頭微微推近，光線映照出油脂豐厚的crema細節。背景柔焦、色彩溫潤，整體呈現安靜、專注、極致質感的瞬間。
This prompt tests two of the hardest elements in AI video generation: liquid and steam.
Sora 2 again delivered the most accurate result. The group head, portafilter, and crema all looked realistic and well-rendered. The only small flaw was the last drips of espresso, which felt slightly unnatural.
Jimeng produced a decent result but had a few noticeable issues. The portafilter looked oversized, and the model unexpectedly added latte art to the cup. Still, the steam effects were surprisingly good.
Wan 2.5’s result stood out for its slow, smooth camera movement — a nice touch. Unfortunately, the machine looked overly shiny and metallic. The biggest issue was that the cup started completely full and never overflowed despite the espresso continuing to drip.
Ranking: Sora 2 > Jimeng AI > Wan 2.5
Prompt 3
明亮早晨室內光線下，一名家庭主婦站在落地玻璃窗前（臉部不需特寫，以自然日常姿態呈現）。她握著透明噴瓶朝玻璃窗表面均勻噴灑清潔液（噴霧化成細緻氣霧，微小水珠在玻璃表面迅速散開，形成薄薄的水膜）。鏡頭切至更近距離，玻璃表面反射著室內柔光（可看到水珠滑落的軌跡、細微光線折射）。她以乾淨的白色超細纖維布開始擦拭（布料微微蓬鬆、邊緣略有纖維細節），手部動作輕柔且有節奏。擦拭時玻璃上的霧面區逐漸變得透亮（清晰反射出外面的藍天與庭院輪廓）。最後畫面呈現光滑無痕的玻璃表面（表面反光自然、乾淨無水痕），整體氛圍溫暖、清潔、日常生活質感。
This is the most complex prompt, combining liquid, glass, reflections, textures, microfiber cloth details, lighting, and a short narrative. Glass and water reflections are notoriously difficult for video AIs.
All three models performed surprisingly well with the glass surface and reflections. Wan 2.5 — which struggled in the first two tests — delivered a strong output here. The spray, water flow, reflections, and cleaning motion were all convincing.
Jimeng’s video leaned unintentionally comedic. The spray particles were too large, the character cleaned with exaggerated movements, and she randomly held two bottles of cleaner for no clear reason.
Sora 2 handled the scene elegantly. Like the lemon-cutting test, it not only understood the physics but also added a natural sense of storytelling. While Wan 2.5’s version felt like a casual home video, Sora 2’s output looked more like a product commercial.
Ranking: Sora 2 = Wan 2.5 > Jimeng AI
Verdict
Sora 2 is the clear overall winner. It consistently produced the most realistic physics, strongest visual logic, and most polished storytelling across all three prompts. Wan 2.5 showed flashes of potential — especially in handling glass and reflections — but struggled with basic motion and object interactions in simpler tests. Jimeng AI generated surprisingly natural human characters but lacked consistency and often drifted into illogical or exaggerated behavior.
For creators in Hong Kong looking for the most reliable and high-quality AI video generator available today, Sora 2 still leads the field. That said, both Wan 2.5 and Jimeng AI are evolving quickly, and given the speed of China’s AI development ecosystem, they’re worth keeping an eye on.
About Us
Based in Hong Kong, JoJo Ventures is a specialized production studio blending years of cinematic expertise with the power of CGI and AI. As the AI wave transforms the creative industry, we help businesses break through traditional production bottlenecks. Our mission is to provide more efficient, creative, and scalable ways for companies to communicate their vision.
Our work is trusted by global giants and local icons alike, including
- Pfizer
- Bosch
- Siemens
- Wellcome
- Eu Yan Sang
- SaSa
From premium commercials to the next generation of AI-generated visuals, we are your partners in the AI era.
Let’s build the future of your brand.
📧 Email: business@jojo.ventures
📱 WhatsApp: +852 9853 7469